Daily Papers

Epilepsy is a common disease of the nervous system. Timely prediction of seizures and intervention treatment can significantly reduce the accidental injury of patients and protect the life and health of patients. This paper presents a neuromorphic Spiking Convolutional Transformer, named Spiking Conformer, to detect and predict epileptic seizure segments from scalped long-term electroencephalogram (EEG) recordings. We report evaluation results from the Spiking Conformer model using the Boston Children's Hospital-MIT (CHB-MIT) EEG dataset. By leveraging spike-based addition operations, the Spiking Conformer significantly reduces the classification computational cost compared to the non-spiking model. Additionally, we introduce an approximate spiking neuron layer to further reduce spike-triggered neuron updates by nearly 38% without sacrificing accuracy. Using raw EEG data as input, the proposed Spiking Conformer achieved an average sensitivity rate of 94.9% and a specificity rate of 99.3% for the seizure detection task, and 96.8%, 89.5% for the seizure prediction task, and needs >10x fewer operations compared to the non-spiking equivalent model.

Accurate detection of pulmonary nodules with high sensitivity and specificity is essential for automatic lung cancer diagnosis from CT scans. Although many deep learning-based algorithms make great progress for improving the accuracy of nodule detection, the high false positive rate is still a challenging problem which limited the automatic diagnosis in routine clinical practice. In this paper, we propose a novel pulmonary nodule detection framework based on a 3D Feature Pyramid Network (3DFPN) to improve the sensitivity of nodule detection by employing multi-scale features to increase the resolution of nodules, as well as a parallel top-down path to transit the high-level semantic features to complement low-level general features. Furthermore, a High Sensitivity and Specificity (HS^2) network is introduced to eliminate the falsely detected nodule candidates by tracking the appearance changes in continuous CT slices of each nodule candidate. The proposed framework is evaluated on the public Lung Nodule Analysis (LUNA16) challenge dataset. Our method is able to accurately detect lung nodules at high sensitivity and specificity and achieves 90.4% sensitivity with 1/8 false positive per scan which outperforms the state-of-the-art results 15.6%.

The Mini-SiTian (MST) project is a pathfinder for China's next-generation large-scale time-domain survey, SiTian, aimed at discovering variable stars, transients, and explosive events. MST generates hundreds of thousands of transient alerts every night, approximately 99\% of which are false alarms, posing a significant challenge to its scientific goals. To mitigate the impact of false positives, we propose a deep learning-based solution and systematically evaluate thirteen convolutional neural networks. The results show that ResNet achieves exceptional specificity (99.70\%), EfficientNet achieves the highest recall rate (98.68\%), and DenseNet provides balanced performance with a recall rate of 94.55\% and specificity of 98.66\%. Leveraging these complementary strengths, we developed a bagging-based ensemble classifier that integrates ResNet18, DenseNet121, and EfficientNet\_B0 using a soft voting strategy. This classifier achieved the best AUC value (0.9961) among all models, with a recall rate of 95.37\% and specificity of 99.25\%. It has now been successfully deployed in the MST real-time data processing pipeline. Validation using 5,000 practically processed samples with a classification threshold of 0.798 showed that the classifier achieved 88.31\% accuracy, 91.89\% recall rate, and 99.82\% specificity, confirming its effectiveness and robustness under real application conditions.

Vision-Language Models (VLMs) are increasingly deployed in settings where reliable visual grounding carries operational consequences, yet their behavior under progressively coercive prompt phrasing remains undercharacterized. Existing hallucination benchmarks predominantly rely on neutral prompts and binary detection, leaving open how both the incidence and the intensity of fabrication respond to graded linguistic pressure across structurally distinct task types. We present Ghost-100, a procedurally constructed benchmark of 800 synthetically generated images spanning eight categories across three task families: text-illegibility, time-reading, and object-absence, each designed under a negative-ground-truth principle that guarantees the queried target is absent, illegible, or indeterminate by construction. Every image is paired with five prompts drawn from a structured 5-Level Prompt Intensity Framework, holding the image and task identity fixed while varying only directive force, so that tone is isolated as the sole independent variable. We adopt a dual-track evaluation protocol: a rule-based H-Rate measuring the proportion of responses in which a model crosses from grounded refusal into unsupported positive commitment, and a GPT-4o-mini-judged H-Score on a 1-5 scale characterizing the confidence and specificity of fabrication once it occurs. We additionally release a three-stage automated validation workflow, which retrospectively confirms 717 of 800 images as strictly compliant. Evaluating nine open-weight VLMs, we find that H-Rate and H-Score dissociate substantially across model families, reading-style and presence-detection subsets respond to prompt pressure in qualitatively different ways, and several models exhibit non-monotonic sensitivity peaking at intermediate tone levels: patterns that aggregate metrics obscure.

Some recent studies have described deep convolutional neural networks to diagnose breast cancer in mammograms with similar or even superior performance to that of human experts. One of the best techniques does two transfer learnings: the first uses a model trained on natural images to create a "patch classifier" that categorizes small subimages; the second uses the patch classifier to scan the whole mammogram and create the "single-view whole-image classifier". We propose to make a third transfer learning to obtain a "two-view classifier" to use the two mammographic views: bilateral craniocaudal and mediolateral oblique. We use EfficientNet as the basis of our model. We "end-to-end" train the entire system using CBIS-DDSM dataset. To ensure statistical robustness, we test our system twice using: (a) 5-fold cross validation; and (b) the original training/test division of the dataset. Our technique reached an AUC of 0.9344 using 5-fold cross validation (accuracy, sensitivity and specificity are 85.13% at the equal error rate point of ROC). Using the original dataset division, our technique achieved an AUC of 0.8483, as far as we know the highest reported AUC for this problem, although the subtle differences in the testing conditions of each work do not allow for an accurate comparison. The inference code and model are available at https://github.com/dpetrini/two-views-classifier

Large language models (LLMs) promise to accelerate incident response in production systems, yet single-agent approaches generate vague, unusable recommendations. We present MyAntFarm.ai, a reproducible containerized framework demonstrating that multi-agent orchestration fundamentally transforms LLM-based incident response quality. Through 348 controlled trials comparing single-agent copilot versus multi-agent systems on identical incident scenarios, we find that multi-agent orchestration achieves 100% actionable recommendation rate versus 1.7% for single-agent approaches, an 80 times improvement in action specificity and 140 times improvement in solution correctness. Critically, multi-agent systems exhibit zero quality variance across all trials, enabling production SLA commitments impossible with inconsistent single-agent outputs. Both architectures achieve similar comprehension latency (approx.40s), establishing that the architectural value lies in deterministic quality, not speed. We introduce Decision Quality (DQ), a novel metric capturing validity, specificity, and correctness properties essential for operational deployment that existing LLM metrics do not address. These findings reframe multi-agent orchestration from a performance optimization to a production-readiness requirement for LLM-based incident response. All code, Docker configurations, and trial data are publicly available for reproduction.

Human teaching effort is a significant bottleneck for the broader applicability of interactive imitation learning. To reduce the number of required queries, existing methods employ active learning to query the human teacher only in uncertain, risky, or novel situations. However, during these queries, the novice's planned actions are not utilized despite containing valuable information, such as the novice's capabilities, as well as corresponding uncertainty levels. To this end, we allow the novice to say: "I plan to do this, but I am uncertain." We introduce the Active Skill-level Data Aggregation (ASkDAgger) framework, which leverages teacher feedback on the novice plan in three key ways: (1) S-Aware Gating (SAG): Adjusts the gating threshold to track sensitivity, specificity, or a minimum success rate; (2) Foresight Interactive Experience Replay (FIER), which recasts valid and relabeled novice action plans into demonstrations; and (3) Prioritized Interactive Experience Replay (PIER), which prioritizes replay based on uncertainty, novice success, and demonstration age. Together, these components balance query frequency with failure incidence, reduce the number of required demonstration annotations, improve generalization, and speed up adaptation to changing domains. We validate the effectiveness of ASkDAgger through language-conditioned manipulation tasks in both simulation and real-world environments. Code, data, and videos are available at https://askdagger.github.io.

AI-text detection generally asks whether a machine wrote a text. As AI becomes an increasingly integral and accepted part of writing, the more relevant question is no longer whether AI was used, but how much of the text reflects human contribution. A prominent line of work estimates this by measuring how predictive a person's prompt is of the final text. We show that this measure can be easily manipulated. We introduce SPECIFICITY PADDING that simply asks the model to copy a few random IDs from the prompt, and that alone makes a title-only request look more human than one that supplies the article's key ideas. A context-swap audit then holds each document fixed, and only shows or hides the IDs to the evaluator. The score of the unchanged text swings by a third to half of its scale, and it rises for every document we test, on news and scientific abstracts, across varied LLMs from 3.8B to 753B parameters. The same articles can average 85\% human or 15\% human depending only on which prompt is on record, and even with no prompt at all, the output-only variant rewards the random strings themselves. An exact identity, the credit ledger, explains why. The score is a surprisal-weighted average over the text, so rare strings that the context makes copyable collect the credit, whatever they mean. A two-point bound marks a deeper limit. If the prompt could have been machine-written, every transcript-only method must, in the worst case, miss by half the contribution at stake. Likelihood-based scores alone cannot reliably establish human contribution without evidence of what the person added and how the text was produced.

The false discovery rate (FDR) measures the share of false positives in a set of statistical tests. I develop simple and intuitive bounds on the FDR in cross-sectional predictability publications. The simplest bound requires just a few lines of math and finds FDR le 25% based on summary statistics in eight out of nine previous studies. A more refined bound finds FDR le 9%. The FDR is small because randomly selecting accounting ratios produces statistically significant predictability far more often than would occur if there were no predictability. The bounds also reconcile the disparate FDR estimates in the literature.

Benchmarks like MMLU suggest flagship language models approach factuality saturation above 90\%. LLMpedia shows this picture is incomplete. We materialize {sim}1.3M encyclopedia articles entirely from parametric memory across three model families, then audit every claim against Wikipedia and curated web evidence. For gpt-5-mini, the verifiable true rate is 68.4\% on Wikipedia-covered subjects - more than 21\,pp below MMLU - and the gap is driven by unverifiability (30.5\%), not refutation (1.2\%). Beyond Wikipedia, frontier articles audited against curated web evidence reach 57.6\%; Wikipedia covers only 56.7\% of model-surfaced subjects, and three model families overlap in just 7.3\% of subject choices. In a retrieval-trap benchmark inspired by prior analysis of Grokipedia, LLMpedia is more factual at roughly half the textual similarity to Wikipedia. Every prompt, article, and verdict is released. Data, code, interface: https://llmpedia.net.

We introduce a corpus of 7,032 sentences rated by human annotators for formality, informativeness, and implicature on a 1-7 scale. The corpus was annotated using Amazon Mechanical Turk. Reliability in the obtained judgments was examined by comparing mean ratings across two MTurk experiments, and correlation with pilot annotations (on sentence formality) conducted in a more controlled setting. Despite the subjectivity and inherent difficulty of the annotation task, correlations between mean ratings were quite encouraging, especially on formality and informativeness. We further explored correlation between the three linguistic variables, genre-wise variation of ratings and correlations within genres, compatibility with automatic stylistic scoring, and sentential make-up of a document in terms of style. To date, our corpus is the largest sentence-level annotated corpus released for formality, informativeness, and implicature.

Evaluation budgets in agentic retrieval-augmented generation span questions, search trajectories, and repeated answers. We measure allocation precision, reading efficiency, and cost boundaries using a retrieval-feedback comparison on HotpotQA and MuSiQue. At 34.14--34.39M model tokens, broader question coverage lowers standard error by 33\% versus five reads and 12.6\% versus three trajectories. Archived nested and Q-only forecasts predict these allocations within 4.0\% and 3.5\%, respectively. Depth subsets establish no clear forecasting advantage beyond the two-trajectory audit. One-read variance penalties relative to the fitted optimum at the same token budget are 0--9.9\%, with substantial Pro uncertainty. Under recorded model fees, more questions beat more trajectories at search prices of \$0--1 per 1,000 requests; question-versus-read fee rankings remain unresolved. Temperature zero cuts answer disagreement from 14.3\% to 3.4\% while comparison precision stays similar. \par\medskip\noindentKeywords: Agentic RAG; Evaluation budget; Generalizability theory; Repeated sampling.

AI chatbots are rapidly shaping how people encounter the news, yet no prior study has systematically measured how accurately these systems, with their proprietary search integrations and retrieval-synthesis pipelines, handle emerging facts across languages and regions. We present a 14-day (February 9-22, 2026) evaluation of six AI chatbots (Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini) on 2,100 factual questions derived from same-day BBC News reporting across six regional services (US & Canada, Arabic, Afrique, Hindi, Russian, Turkish). The best systems achieve over 90% multiple-choice accuracy on questions about events reported hours earlier. The same systems, however, lose 11-13% under free-response evaluation, and 16-17% across the cohort. We further characterize three failure patterns. First, every model achieves its lowest accuracy on Hindi (79% vs. 89-91% elsewhere) and citations indicate an Anglophone retrieval bias (e.g., models answering Hindi queries cite English Wikipedia more than any Hindi outlet). Second, retrieval, not reasoning, failures drive over 70% of all errors. When models retrieve a correct source, they often extract the correct answer; the problem is to land on the right source in the first place. Third, models achieving 88-96% accuracy on well-formed questions drop to 19-70% when questions contain subtle false premises, with the most vulnerable model accepting fabricated facts 64% of the time. We also identify a detection-accuracy paradox: the best false-premise detector ranks second in adversarial accuracy (abstention rate), while a weaker detector ranks first, showing that premise detection and answer recovery are partially independent capabilities. Overall, these suggest that high accuracy can mask systematic regional inequity, near-total dependence on retrieval infrastructure, and vulnerability to imperfect queries real users pose.

Instruction tuning is now the default way to train and adapt large language models, but many instruction--input--output pairs are only weakly specified: for a given input, the same output can remain plausible under several alternative instructions. This raises a simple question: does the instruction uniquely determine the target output? We propose the Task--Specificity Score (TSS) to quantify how much an instruction matters for predicting its output, by contrasting the true instruction against plausible alternatives for the same input. We further introduce TSS++, which uses hard alternatives and a small quality term to mitigate easy-negative effects. Across three instruction datasets (Alpaca, Dolly-15k, NI-20) and three open LLMs (Gemma, Llama, Qwen), we show that selecting task-specific examples improves downstream performance under tight token budgets and complements quality-based filters such as perplexity and IFD.

The proliferation of large language models has raised growing concerns about their misuse, particularly in cases where AI-generated text is falsely attributed to human authors. Machine-generated content detectors claim to effectively identify such text under various conditions and from any language model. This paper critically evaluates these claims by assessing several popular detectors (RADAR, Wild, T5Sentinel, Fast-DetectGPT, PHD, LogRank, Binoculars) on a range of domains, datasets, and models that these detectors have not previously encountered. We employ various prompting strategies to simulate practical adversarial attacks, demonstrating that even moderate efforts can significantly evade detection. We emphasize the importance of the true positive rate at a specific false positive rate (TPR@FPR) metric and demonstrate that these detectors perform poorly in certain settings, with TPR@.01 as low as 0%. Our findings suggest that both trained and zero-shot detectors struggle to maintain high sensitivity while achieving a reasonable true positive rate.

Measuring complexity from a model's own code is biased by failure: a hard prompt can produce a short, broken program that looks simple. We instead score each prompt's structure on six dimensions from the prompt alone, never from generated code or test results. We draw 5,000 Python prompts across six bands of a preliminary single-rater score, and four out-of-panel LLM judges rescore every prompt (composite ICC(2,1) = 0.872). Twenty-one models answer every prompt (105,000 generations), and an independent LLM auditor, checked against known-answer cases, reviews each unit-test verdict. Averaged over models, pass rate has a nonmonotone breakpoint at composite 14.0: 83.4% at or below, 90.1% above. That jump is not a law of difficulty: with task-type fixed effects the breakpoint stays at 14.0 but the adjusted jump shrinks from 6.8 to 0.4 points, and neither of the benchmark's two prompt sources reproduces it on its own. Nineteen models pass more often above their own breakpoint and two less often; one of those declines comes from API calls that returned nothing. In post hoc follow-up, Kimi K2.5 shows the clearest decline, at 14.25, under a prompt-source control and within the later source alone. Among failed generations with measurable output complexity, 25.0% pair a prompt composite above 8 with output complexity of at most 10, twice the share among passing ones. The pooled breakpoint moves to 11.25 on raw test outcomes. Human agreement with the index is moderate, paraphrase and cross-language rescoring preserve its ordering, and overidentification tests reject the six dimensions as instruments, so we make no causal claim. The contribution is a prompt-side way to locate where each model's pass rate changes and to test how much that depends on the prompt mix and on how success is scored.

A unique aspect of ColBERT is its use of [MASK] tokens in queries to score documents (query augmentation). Prior work shows [MASK] tokens weighting non-[MASK] query terms, emphasizing certain tokens over others , rather than introducing whole new terms as initially proposed. We begin by demonstrating that a term weighting behavior previously reported for [MASK] tokens in ColBERTv1 holds for ColBERTv2. We then examine the effect of changing the number of [MASK] tokens from zero to up to four times past the query input length used in training, both for first stage retrieval, and for scoring candidates, observing an initial decrease in performance with few [MASK]s, a large increase when enough [MASK]s are added to pad queries to an average length of 32, then a plateau in performance afterwards. Additionally, we compare baseline performance to performance when the query length is extended to 128 tokens, and find that differences are small (e.g., within 1% on various metrics) and generally statistically insignificant, indicating performance does not collapse if ColBERT is presented with more [MASK] tokens than expected.

Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs. In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes? To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates. We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases. When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts. Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.

Leading vision-language models (VLMs) are trained on general Internet content, overlooking scientific journals' rich, domain-specific knowledge. Training on specialty-specific literature could yield high-performance, task-specific tools, enabling generative AI to match generalist models in specialty publishing, educational, and clinical tasks. We created NeuroPubs, a multimodal dataset of 23,000 Neurosurgery Publications articles (134M words, 78K image-caption pairs). Using NeuroPubs, VLMs generated publication-ready graphical abstracts (70% of 100 abstracts) and board-style questions indistinguishable from human-written ones (54% of 89,587 questions). We used these questions to train CNS-Obsidian, a 34B-parameter VLM. In a blinded, randomized controlled trial, our model demonstrated non-inferiority to then state-of-the-art GPT-4o in neurosurgical differential diagnosis (clinical utility, 40.62% upvotes vs. 57.89%, p=0.1150; accuracy, 59.38% vs. 65.79%, p=0.3797). Our pilot study demonstrates how training generative AI models on specialty-specific journal content - without large-scale internet data - results in high-performance academic and clinical tools, enabling domain-tailored AI across diverse fields.

Counterfactual fairness audits of clinical language-model agents report a flip rate: how often an action changes when only the patient's demographic descriptor changes. Part of that rate is not demographic. A stochastic agent also changes its own action when nothing changes, and a flip rate cannot be interpreted without knowing how often. We measured it. Re-running one condition ten times over sixteen synthetic vignettes at default sampling changed a clinical agent's action in 8.7 percent of replicate pairs, from 2.2 percent for intensive-care escalation to 17.9 percent for controlled-substance caution, an output given no operational criteria. Across six models from five vendors, pooled floors ranged from 2.5 to 23.7 percent; in this panel neither disclosed size, vendor, nor hosting ordered them. The floor depends on the decoding configuration: majority voting over five draws removed 39 percent of it (95 percent confidence interval (CI) 18 to 64); at temperature 0 three of four locally served models showed no disagreement, but a hosted model still did. We also show that, for a binary action, the flip rate expected under no demographic effect equals the floor and a real effect adds only its square, so a flip rate inside the floor is not evidence of fairness, and direction must be tested with a signed paired test. We give a four-step reporting procedure and release FairMedAgent, the harness, with its protocol, vignettes and analysis scripts, so any team can measure the floor for its own agent.

Ambient AI scribes draft clinical notes under the reassurance that a clinician signs every note. We audited three commercial AI scribes on the same 142 consultations: 565 notes from recorded UK primary-care and US ambulatory encounters plus authored scenarios. Twelve discovery passes proposed 13,678 candidate errors; the 5,898 clearing an importance filter went to an adversarial panel of two models from different families, each told to refute what it could, and 618 survived. One note in three (31.3% [27.0, 35.6]) carries a verified failure, concentrated in allergy and medication information, invented patient identity, and history written up as examination on telephone consultations that can contain none. No product was given a patient record; setting aside the two classes a record would have prefilled, invented identity and dates, the rate is 24.8% [20.8, 29.0]. One failure mode did not fit our scheme, drawn from published scribe-error taxonomies: a treatment the clinician retracts, recorded as delivered care. Two clinicians adjudicated blind, disjoint samples: a physician author upheld 20 of 21 findings (95.2% [77.3, 99.2]) and an independent clinician, not an author, 12 of 12 ([75.8, 100]); both judged every sampled refusal genuine. A failure rate depends on the instrument as much as the scribes. With model, evidence and settings fixed, the review instruction alone moves the share of candidates verified from 9.3% to 79.0%, and the reviewing family moves it too: alone at that instruction the gentler flags 54.8% of notes against 27.8%. Between 28% and 97% of sampled notes carry a failure depending on the standard. Published audits disagree among themselves by a margin instrument differences alone can produce: omission is 54-86% of their errors against our 23.1%. We release all 618 findings with transcript-side evidence, every prompt and model version, and the re-runnable pipeline.

This paper describes a system designed to distinguish between AI-generated and human-written scientific excerpts in the DAGPap24 competition hosted within the Fourth Workshop on Scientific Document Processing. In this competition the task is to find artificially generated token-level text fragments in documents of a scientific domain. Our work focuses on the use of a multi-task learning architecture with two heads. The application of this approach is justified by the specificity of the task, where class spans are continuous over several hundred characters. We considered different encoder variations to obtain a state vector for each token in the sequence, as well as a variation in splitting fragments into tokens to further feed into the input of a transform-based encoder. This approach allows us to achieve a 9% quality improvement relative to the baseline solution score on the development set (from 0.86 to 0.95) using the average macro F1-score, as well as a score of 0.96 on a closed test part of the dataset from the competition.

Prior work on a social web highlighter located individuality in selection -- which documents a person chooses to highlight -- but measured it cross-sectionally. We ask the temporal question: is a reader's selection signature a trait or a state? We freeze each reader's first six months of highlighting as a profile and track its own-vs-other advantage on their later selections at growing gaps (to 24+ months), with negatives drawn from the same calendar era -- so supply drift cannot masquerade as personal drift -- at a coarse global level and at a fine level whose negatives and controls come from the reader's own interest neighborhood; the anchor cell reproduces the prior cross-sectional level (+0.188 vs +0.169), validating the harness. Four results. Within the same users, the fine-layer advantage shows no statistically detectable paired decline at any horizon (6-12 month retention R = 1.00 [0.85, 1.18], n = 212; the farthest bin is compatible with a modest decline; the only contrast whose interval excludes zero is the coarse layer at 12-24 months, about 13%). The signal is not reducible to repeated domains (~90% survives excluding all profile sources). Within-person drift is slow (a recent-half profile beats the old half by +0.042). Prospectively, personal profiles -- even one built from a reader's earliest documents, median 20 months before evaluation -- rank their next reads at roughly 3x the AP of every simple non-personal prior tested. We use "trait" operationally (a stable signature under continued engagement); the scope is heavy, long-tenured readers of one platform, and exposure is not separable from choice.

Character Error Rate (CER), the standard metric for OCR and vision-language models (VLMs), charges a typo as much as a wrong name. We introduce semCER, which keeps CER's alignment and denominator and charges each grapheme edit a severity from an LLM judge, with no word segmentation. Normalized CER keeps the score of a whole transcription, where it tracks readers as well as semCER or better, and semCER breaks that score down into named edits. The charges track readers' ratings edit by edit in English, Vietnamese, and Chinese and put the edits readers rate most severe near the top of a sorted list. On pairs drawn where CER and semCER disagreed, readers side with semCER more often than with normalized CER in English and Chinese, and equally often in Vietnamese. Across 16 languages, the judge charges a wrong diacritic or tone mark almost nothing in Spanish but nearly a wrong letter's severity in Telugu. Where we can count real words, a general VLM's wrong letter forms one far above chance while a classical pipeline's stays near it. We release the 61 million judged edits, the aggregated reader ratings, and the distilled scorer at https://github.com/nguyenx2t/semcer.

Large language models (LLMs) have demonstrated unparalleled prowess in mimicking human-like text generation and processing. Among the myriad of applications that benefit from LLMs, automated code generation is increasingly promising. The potential to transform natural language prompts into executable code promises a major shift in software development practices and paves the way for significant reductions in manual coding efforts and the likelihood of human-induced errors. This paper reports the results of a study that evaluates the performance of various LLMs, such as Bard, ChatGPT-3.5, ChatGPT-4, and Claude-2, in generating Python for coding problems. We focus on how levels of prompt specificity impact the accuracy, time efficiency, and space efficiency of the generated code. A benchmark of 104 coding problems, each with four types of prompts with varying degrees of tests and specificity, was employed to examine these aspects comprehensively. Our results indicate significant variations in performance across different LLMs and prompt types, and its key contribution is to reveal the ideal prompting strategy for creating accurate Python functions. This study lays the groundwork for further research in LLM capabilities and suggests practical implications for utilizing LLMs in automated code generation tasks and test-driven development.

Typed-decision models score options that the caller supplies at request time, so their accuracy depends on label wording that the model's developer does not control. We compare two releases of a 307M-parameter multilingual cross-encoder, Ru-Jev, trained on thirteen Russian-language tasks; the later release adds an augmentation that presents labels in several surface forms. Holding questions fixed and varying only how eighteen routing labels are worded, we evaluate ten label vocabularies, produced by three separate processes and checked for overlap with the training data, with five candidates per question and, for eight of them, also with eighteen. On held-out vocabularies made of ordinary or borrowed words (English-derived jargon, office slang, department names), the augmented release is more accurate in every setting, by up to 0.24. Eleven of these fourteen gains survive a Holm correction (slang with eighteen candidates and descriptive sentences with five do not), and for jargon on all labels the label-level interval excludes zero in both settings. On abbreviations and opaque codes it gives no usable gain: on abbreviations it is not better, and worse on the clean slice, and on opaque codes both releases stay at or below chance once labels related to training options are removed. In most cells where accuracy rises, the risk-coverage area falls, but with eighteen candidates confident errors rise on jargon and slang. On two datasets neither release saw, accuracy moves by -8.4 and +5.6 points. We also report four evaluation hazards from comparing the two releases.

The proliferation of Large Language Models (LLMs) highlights the critical importance of conducting thorough evaluations to discern their comparative advantages, limitations, and optimal use cases. Particularly important is assessing their capacity to accurately retrieve information included in a given prompt. A model's ability to do this significantly influences how effectively it can utilize contextual details, thus impacting its practical efficacy and dependability in real-world applications. Our research analyzes the in-context recall performance of various LLMs using the needle-in-a-haystack method. In this approach, a factoid (the "needle") is embedded within a block of filler text (the "haystack"), which the model is asked to retrieve. We assess the recall performance of each model across various haystack lengths and with varying needle placements to identify performance patterns. This study demonstrates that an LLM's recall capability is not only contingent upon the prompt's content but also may be compromised by biases in its training data. Conversely, adjustments to model architecture, training strategy, or fine-tuning can improve performance. Our analysis provides insight into LLM behavior, offering direction for the development of more effective applications of LLMs.

We introduce MURA, a large dataset of musculoskeletal radiographs containing 40,561 images from 14,863 studies, where each study is manually labeled by radiologists as either normal or abnormal. To evaluate models robustly and to get an estimate of radiologist performance, we collect additional labels from six board-certified Stanford radiologists on the test set, consisting of 207 musculoskeletal studies. On this test set, the majority vote of a group of three radiologists serves as gold standard. We train a 169-layer DenseNet baseline model to detect and localize abnormalities. Our model achieves an AUROC of 0.929, with an operating point of 0.815 sensitivity and 0.887 specificity. We compare our model and radiologists on the Cohen's kappa statistic, which expresses the agreement of our model and of each radiologist with the gold standard. Model performance is comparable to the best radiologist performance in detecting abnormalities on finger and wrist studies. However, model performance is lower than best radiologist performance in detecting abnormalities on elbow, forearm, hand, humerus, and shoulder studies. We believe that the task is a good challenge for future research. To encourage advances, we have made our dataset freely available at https://stanfordmlgroup.github.io/competitions/mura .

This study explores the relationship between textual features and Information Engagement (IE) on digital platforms. It highlights the impact of computational linguistics and analytics on user interaction. The READ model is introduced to quantify key predictors like representativeness, ease of use, affect, and distribution, which forecast engagement levels. The model's effectiveness is validated through AB testing and randomized trials, showing strong predictive performance in participation (accuracy: 0.94), perception (accuracy: 0.85), perseverance (accuracy: 0.81), and overall IE (accuracy: 0.97). While participation metrics are strong, perception and perseverance show slightly lower recall and F1-scores, indicating some challenges. The study demonstrates that modifying text based on the READ model's insights leads to significant improvements. For example, increasing representativeness and positive affect boosts selection rates by 11 percent, raises evaluation averages from 3.98 to 4.46, and improves retention rates by 11 percent. These findings highlight the importance of linguistic factors in IE, providing a framework for enhancing digital text engagement. The research offers practical strategies applicable to fields like education, health, and media.

Medical coding translates clinical documentation into standardized codes for billing, research, and public health, but manual coding is time-consuming and error-prone. Existing automation efforts rely on small datasets that poorly represent real-world patient heterogeneity. We trained a language model on 5.8 million electronic health records from 1.8 million patients across nearly all specialties in Eastern Denmark (2006--2016) to predict ICD-10 codes from clinical notes, medications, and laboratory results. Evaluated on 270,000 held-out patients, the model achieved a micro F1 of 71.8% and a top-10 recall of 95.5%. Performance varied by specialty (F1: 53--91%), with higher scores in specialties with well-defined diagnostic criteria. Codes appearing predominantly as secondary diagnoses had markedly lower F1 scores. For three such codes (suicide-related behaviors, weight disorders, and hypertension), the model identified thousands of uncoded cases, of which 76-86% were confirmed valid upon manual review, suggesting systematic under-coding rather than model error. These findings suggest under-coding of secondary diagnoses in Eastern Denmark during this period, with potential implications for epidemiological research, public health surveillance, and understanding of multimorbidity. Similar time constraints and reimbursement structures in other healthcare systems suggest this may not be isolated to this dataset. The model can automate coding for approximately 50% of cases and provide accurate suggestions for most others, and may offer a practical solution to help capture missed secondary conditions.

With the surge of large language models (LLMs) and their ability to produce customized output, style-personalized text generation--"write like me"--has become a rapidly growing area of interest. However, style personalization is highly specific, relative to every user, and depends strongly on the pragmatic context, which makes it uniquely challenging. Although prior research has introduced benchmarks and metrics for this area, they tend to be non-standardized and have known limitations (e.g., poor correlation with human subjects). LLMs have been found to not capture author-specific style well, it follows that the metrics themselves must be scrutinized carefully. In this work we critically examine the effectiveness of the most common metrics used in the field, such as BLEU, embeddings, and LLMs-as-judges. We evaluate these metrics using our proposed style discrimination benchmark, which spans eight diverse writing tasks across three evaluation settings: domain discrimination, authorship attribution, and LLM-generated personalized vs non-personalized discrimination. We find strong evidence that employing ensembles of diverse evaluation metrics consistently outperforms single-evaluator methods, and conclude by providing guidance on how to reliably assess style-personalized text generation.

The potential of large language models (LLMs) to simultaneously perform a wide range of natural language processing (NLP) tasks has been the subject of extensive research. Although instruction tuning has proven to be a data-efficient method for transforming LLMs into such generalist models, their performance still lags behind specialist models trained exclusively for specific tasks. In this paper, we investigate whether incorporating broad-coverage generalist instruction tuning can contribute to building a specialist model. We hypothesize that its efficacy depends on task specificity and skill requirements. Our experiments assess four target tasks with distinct coverage levels, revealing that integrating generalist instruction tuning consistently enhances model performance when the task coverage is broad. The effect is particularly pronounced when the amount of task-specific training data is limited. Further investigation into three target tasks focusing on different capabilities demonstrates that generalist instruction tuning improves understanding and reasoning abilities. However, for tasks requiring factual knowledge, generalist data containing hallucinatory information may negatively affect the model's performance. Overall, our work provides a systematic guide for developing specialist models with general instruction tuning. Our code and other related resources can be found at https://github.com/DavidFanzz/Generalist_or_Specialist.

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

Large language models (LLMs) now write code in settings where misreading a single word can break safety or cost money, yet we still expect them to overlook stray typos. To probe where useful robustness ends and harmful insensitivity begins, we compile 50 LeetCode problems and craft three minimal prompt perturbations that should vary in importance: (i) progressive underspecification deleting 10 % of words per step; (ii) lexical flip swapping a pivotal quantifier ("max" to "min"); and (iii) jargon inflation replacing a common noun with an obscure technical synonym. Six frontier models, including three "reasoning-tuned" versions, solve each mutated prompt, and their Python outputs are checked against the original test suites to reveal whether they reused the baseline solution or adapted. Among 11 853 generations we observe a sharp double asymmetry. Models remain correct in 85 % of cases even after 90 % of the prompt is missing, showing over-robustness to underspecification, yet only 54 % react to a single quantifier flip that reverses the task, with reasoning-tuned variants even less sensitive than their bases. Jargon edits lie in between, passing through 56 %. Current LLMs thus blur the line between harmless noise and meaning - changing edits, often treating both as ignorable. Masking salient anchors such as function names can force re - evaluation. We advocate evaluation and training protocols that reward differential sensitivity: stay steady under benign noise but adapt - or refuse - when semantics truly change.

Statutes are increasingly parsed by machines before people read them, and the parsers disagree: on Missouri's statutes, two independently written extractors diverge on numeric-threshold presence at a false-negative rate of 0.43. We ask what formal logic survives such noise. We build a passive survival certificate for the Duquenne-Guigues implication basis of machine-extracted statutory contexts: per-attribute inter-extractor disagreement is measured, replayed against the basis in 1,000 Monte Carlo trials, and an implication is certified only when a one-sided Wilson 95% lower bound on survival reaches 0.95; every certified implication carries premise spans and a minimal counterexample. On 29,365 Missouri sections and 502 Indian central-Act sections, the preregistered held-out gate passes (10 statute families across 7 Titles exact; 16 across 11 with 5% tolerance), yet under one globally deployed error model 93.2% of held-out chapters fall below the informativeness floor, and a 2x2 factorial assigns that to calibration-rate transfer, not selection. The certificate is usable but fragile: deploy it per-chapter-calibrated or error-tolerant. Code, data products, and the audit trail, including one retracted claim, are released.

Whether a language model looks demographically biased can depend on how the audit asks its question. A charitable-aid benchmark reports that the same models favor minority applicants when rating requests one at a time and penalize some when ranking side by side. We test whether that reversal generalizes to hiring, lending, and medical triage: 40,726 requests to five models, applications differing only in the applicant's name, and a primary test fixed before collection. It does not. None of 36 planned contrasts survives correction. The rating advantage keeps its sign at roughly half the published size, and a precision extension bounds any hiring ranking penalty below the published effect, though the lending and triage ranking floors sit above that margin, so the exclusion is conclusive for hiring ranking and for rating in all three domains only. Planted disparities tracking their injected sizes and a directional replication on the original aid materials bound these nulls. The audit is livelier than the demographics: models recognize transparent audits nearly always, tie every identical-content comparison whether the varying detail is race or a hobby, and reward first-listed candidates as much as any demographic effect we measure. Audit verdicts reflect audit construction more than demographic bias.

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

We present an empirical evaluation of various outputs generated by nine of the most widely-available large language models (LLMs). Our analysis is done with off-the-shelf, readily-available tools. We find a correlation between percentage of memorized text, percentage of unique text, and overall output quality, when measured with respect to output pathologies such as counterfactual and logically-flawed statements, and general failures like not staying on topic. Overall, 80.0% of the outputs evaluated contained memorized data, but outputs containing the most memorized content were also more likely to be considered of high quality. We discuss and evaluate mitigation strategies, showing that, in the models evaluated, the rate of memorized text being output is reduced. We conclude with a discussion on potential implications around what it means to learn, to memorize, and to evaluate quality text.

Summarization models often generate text that is poorly calibrated to quality metrics because they are trained to maximize the likelihood of a single reference (MLE). To address this, recent work has added a calibration step, which exposes a model to its own ranked outputs to improve relevance or, in a separate line of work, contrasts positive and negative sets to improve faithfulness. While effective, much of this work has focused on how to generate and optimize these sets. Less is known about why one setup is more effective than another. In this work, we uncover the underlying characteristics of effective sets. For each training instance, we form a large, diverse pool of candidates and systematically vary the subsets used for calibration fine-tuning. Each selection strategy targets distinct aspects of the sets, such as lexical diversity or the size of the gap between positive and negatives. On three diverse scientific long-form summarization datasets (spanning biomedical, clinical, and chemical domains), we find, among others, that faithfulness calibration is optimal when the negative sets are extractive and more likely to be generated, whereas for relevance calibration, the metric margin between candidates should be maximized and surprise--the disagreement between model and metric defined candidate rankings--minimized. Code to create, select, and optimize calibration sets is available at https://github.com/griff4692/calibrating-summaries

Online health information seeking is shifting from keyword search, where users consider a ranked list of links, to conversational systems that compose a single answer and curate its citations. Source evaluation therefore passes from user to platform, yet what these systems surface is poorly characterized. We audited three free consumer products (ChatGPT, Perplexity, Google AI Overview) on twenty English mental health questions under two prompt conditions, with a subset of three also translated into six further languages of varying resource tiers. We recorded 15,942 citations across 1,140 responses and 1,713 unique domains, then classified every citation with a nine-category organizational typology applied by a deterministic classifier validated against human coding. Citations were heavily concentrated: the ten most-cited domains accounted for 43.6% of English citations, and government, commercial health, and academic sources were closely matched at roughly 22% each. Platforms differed little in typical citation volume but sharply in consistency and in the source types they favored. Explicitly requesting sources shifted composition only modestly. Non-English queries surfaced fewer citations and were routed to language-appropriate resources at significantly lower rates. We release the typology, classifier, and annotated corpus as reusable instruments for auditing generative health search.

A social highlighter's most useful signal -- which passages a crowd of readers marks -- exists only for documents people have already read. Can the aggregate crowd salience of a document be predicted from its text before its marks accumulate? Prior work on this data found that zero-shot language models recover highlight locations worse than a trivial lead (position) baseline, so we ask whether a model trained on the highlight corpus can beat that baseline. Using a pre-registered ladder of models and a by-document cluster bootstrap, we find a small but robust edge: a logistic ranker over sentence embeddings and positional/contextual features beats the lead baseline by +0.044 average precision (95% CI [+0.029, +0.058]; clears a pre-registered margin delta=0.03 in 97% of resamples, and stable across pipeline re-runs). Two unsupervised extractive baselines (centroid, LexRank-style centrality) lose to lead, and the trained model beats them by +0.108, so the edge is not recovered by generic unsupervised proxies -- it reflects learning from real reader marks. In product terms, precision@3 rises from 0.25 to 0.39 (+55% relative) and the model beats lead on 69% of documents. An ablation attributes the edge to the raw embedding (+0.014) and training augmentation (+0.010), each with a positive CI. The edge is not a temporal-generalization failure, and we find no evidence that content drift or near-duplicate leakage explains it. A standardized regression shows the advantage is governed mainly by document popularity (lower popularity, larger edge) and by label reliability. It nearly vanishes only on the most popular content; there it is the lead baseline that strengthens, not the model that weakens. Because our evaluation conditions on documents that eventually accumulated readers, these results are a retrospective cold-start simulation.

A research proposal can describe an established mechanism correctly while claiming to introduce it. We study whether providing the earlier paper corrects such contribution claims. Three controlled experiments compare proposals generated with a contribution-bearing prior and a same-topic control. Providing the prior yields no clear aggregate reduction in unsupported novelty. Human analysis of 175 interpretable exposed proposals finds that 137 recognize the prior's relevance, but 61 correctly attribute the established contribution. Of 71 proposed remaining distinctions, 37 are covered by the same prior. We introduce Research Idea Novelty Inspection (RINI), which audits contribution claims against evidence, checks the remaining distinction, and applies local revisions. Five human annotators evaluate 1,080 original-revision pairs across three methods. On the same 240 originals judged to require correction, successful repair is 11.7% for Self-Revision, 39.1% for Retrieve-and-Revise, and 72.2% for RINI, with research tasks weighted equally. The improvement over same-evidence direct revision is 33.0 percentage points. The revised proposals retain their research questions and technical methods. These results motivate explicit contribution attribution when using literature to generate and revise research proposals.

State-of-the-art machine learning methods exhibit limited compositional generalization. At the same time, there is a lack of realistic benchmarks that comprehensively measure this ability, which makes it challenging to find and evaluate improvements. We introduce a novel method to systematically construct such benchmarks by maximizing compound divergence while guaranteeing a small atom divergence between train and test sets, and we quantitatively compare this method to other approaches for creating compositional generalization benchmarks. We present a large and realistic natural language question answering dataset that is constructed according to this method, and we use it to analyze the compositional generalization ability of three machine learning architectures. We find that they fail to generalize compositionally and that there is a surprisingly strong negative correlation between compound divergence and accuracy. We also demonstrate how our method can be used to create new compositionality benchmarks on top of the existing SCAN dataset, which confirms these findings.

In natural language processing, a recently popular line of work explores how to best report the experimental results of neural networks. One exemplar publication, titled "Show Your Work: Improved Reporting of Experimental Results," advocates for reporting the expected validation effectiveness of the best-tuned model, with respect to the computational budget. In the present work, we critically examine this paper. As far as statistical generalizability is concerned, we find unspoken pitfalls and caveats with this approach. We analytically show that their estimator is biased and uses error-prone assumptions. We find that the estimator favors negative errors and yields poor bootstrapped confidence intervals. We derive an unbiased alternative and bolster our claims with empirical evidence from statistical simulation. Our codebase is at http://github.com/castorini/meanmax.

We study retrieving a set of documents that covers various perspectives on a complex and contentious question (e.g., will ChatGPT do more harm than good?). We curate a Benchmark for Retrieval Diversity for Subjective questions (BERDS), where each example consists of a question and diverse perspectives associated with the question, sourced from survey questions and debate websites. On this data, retrievers paired with a corpus are evaluated to surface a document set that contains diverse perspectives. Our framing diverges from most retrieval tasks in that document relevancy cannot be decided by simple string matches to references. Instead, we build a language model based automatic evaluator that decides whether each retrieved document contains a perspective. This allows us to evaluate the performance of three different types of corpus (Wikipedia, web snapshot, and corpus constructed on the fly with retrieved pages from the search engine) paired with retrievers. Retrieving diverse documents remains challenging, with the outputs from existing retrievers covering all perspectives on only 33.74% of the examples. We further study the impact of query expansion and diversity-focused reranking approaches and analyze retriever sycophancy. Together, we lay the foundation for future studies in retrieval diversity handling complex queries.

Supervised AI-text detectors report high benchmark accuracy, but it is not clear what their decisions are based on. We analyze a RoBERTa-based detector under semantic, structural, and tokenizer-level perturbations, using the M4 dataset (N = 10,000) and controlled generations (N = 300). When Mistral-7B-Instruct was asked to make machine text sound more human, Verb Diversity rose from 0.77 to 0.92 and the outputs became easier to detect. Detection scores appear to track statistical complexity, which also leads to a 76.3% false-positive rate on formal human writing. As a control, we evaluate event-based Latent Space detection. Paraphrasing changed 87% of its event sequences (Jaccard = 0.067), and homoglyphs altered 70% of the extracted verbs even though extraction still ran (Jaccard = 0.30). Its best domain AUC was 0.577. RoBERTa's robustness seems specific to the features it uses, and structural abstraction did not make detection more robust.

This article presents an experiment in fine-tuning a pretrained causal language model (Meta's Llama 3.1 8B Instruct) for aiding in three fundamental tasks of philological research: chronological and geographic attribution as well as text restoration in ancient Greek inscriptions and documentary papyri. Using a prompt-based instruct approach, the fine-tuned models surpass the state of the art in key metrics. For inscriptions, the models achieve a lower average character error rate (CER) of 22.5% (vs. 26.3%), while closely matching top-1 accuracy (60.9% vs. 61.8%) and top-20 accuracy (77.5% vs. 78.3%) for sequences up to 10 characters. They also provide a practical advantage by ignoring spaces during reconstruction, aligning better with the scriptio continua typically used in ancient written artifacts. In geographic attribution, the model outperforms previous benchmarks with a top-1 accuracy of 75.0% (vs. 70.8%) and a top-3 accuracy of 83.7% (vs. 82.1%). For dating, it achieves an average deviation of 26.2 years (vs. 29.3) and a median deviation of 1 year (vs. 3) from the actual date range. The models also set new baselines for documentary papyri, with a CER of 16.3%, a top-1 accuracy of 71.3%, and top-20 of 85.0% in text reconstruction; a top-1 accuracy of 66.4% and top-3 of 79.9% in geographic attribution; and, in chronological attribution, a deviation of 21.7 years from the actual termini post/ante quem, with a median deviation of 0 years.

Commonly used evaluation measures including Recall, Precision, F-Measure and Rand Accuracy are biased and should not be used without clear understanding of the biases, and corresponding identification of chance or base case levels of the statistic. Using these measures a system that performs worse in the objective sense of Informedness, can appear to perform better under any of these commonly used measures. We discuss several concepts and measures that reflect the probability that prediction is informed versus chance. Informedness and introduce Markedness as a dual measure for the probability that prediction is marked versus chance. Finally we demonstrate elegant connections between the concepts of Informedness, Markedness, Correlation and Significance as well as their intuitive relationships with Recall and Precision, and outline the extension from the dichotomous case to the general multi-class case.

Collecting high-quality training examples for language model fine-tuning is expensive, with practical budgets limiting the amount of data that can be procured. We investigate whether example difficulty affects GRPO training effectiveness by comparing selection strategies (easy, medium, hard, random) across multiple models and reasoning tasks. Training on the hardest 10\% of examples (those where the base model fails most often) yields dramatic performance gains up to 47\%, while easy examples produce minimal improvements of 3-15\%. This occurs because GRPO requires outcome variance to generate learning signals; hard examples maintain mixed success/failure outcomes throughout training while easy examples quickly converge to consistent success, eliminating learning opportunities. Moreover, models trained on hard examples show superior out-of-distribution generalization, with only hard-trained models achieving meaningful gains on the AIME2025 benchmark. Our findings provide clear guidance: when budget-constrained, prioritize collecting and annotating examples where your base model struggles, as these drive nearly all learning value in GRPO fine-tuning

Instruction-tuned language models (LM) are able to respond to imperative commands, providing a more natural user interface compared to their base counterparts. In this work, we present Promptriever, the first retrieval model able to be prompted like an LM. To train Promptriever, we curate and release a new instance-level instruction training set from MS MARCO, spanning nearly 500k instances. Promptriever not only achieves strong performance on standard retrieval tasks, but also follows instructions. We observe: (1) large gains (reaching SoTA) on following detailed relevance instructions (+14.3 p-MRR / +3.1 nDCG on FollowIR), (2) significantly increased robustness to lexical choices/phrasing in the query+instruction (+12.9 Robustness@10 on InstructIR), and (3) the ability to perform hyperparameter search via prompting to reliably improve retrieval performance (+1.4 average increase on BEIR). Promptriever demonstrates that retrieval models can be controlled with prompts on a per-query basis, setting the stage for future work aligning LM prompting techniques with information retrieval.

Safeguards in deployed LLM services are evaluated by refusal, attack success, and policy violation rates. Those rates characterize how a control performed on the requests it was tested on. A deployment has to answer a different question: how much help with harmful tasks the service still gives an attacker who keeps adapting or finds another way in. We determine what each reported result implies for that question, allowing results from different safeguard families to be compared under one deployment criterion. The evidence requirements are strongly asymmetric. One attack that obtains harmful help from the deployed service suffices to establish that such help remains, and such attacks appear repeatedly in the coded record. Establishing that little remains cannot follow from the safeguard's own numbers alone; it also requires evidence about what the surrounding system still allows after the safeguard performs its local function. Such evidence is supported or derived in only a small minority of the depth-coded claims, and one such claim bounds its scoped residual. A better local score is therefore not, by itself, a stronger claim about the deployment. Safeguard research cannot stop at raising local scores; a gain has to be judged by whether it makes a deployed system any safer.

Writing an ad text that attracts people and persuades them to click or act is essential for the success of search engine advertising. Therefore, ad creators must consider various aspects of advertising appeals (A^3) such as the price, product features, and quality. However, products and services exhibit unique effective A^3 for different industries. In this work, we focus on exploring the effective A^3 for different industries with the aim of assisting the ad creation process. To this end, we created a dataset of advertising appeals and used an existing model that detects various aspects for ad texts. Our experiments demonstrated that different industries have their own effective A^3 and that the identification of the A^3 contributes to the estimation of advertising performance.

Why do language models fail on long inputs? The common answer is length. We show that, for retrieval, length is cheap and competition is not. On five open models (0.5B to 3B parameters, four families), neutral filler costs nothing: accuracy stays at 0.97 to 1.00, up to 32K tokens on Qwen2.5-1.5B. Statements in the format of the evidence, placed at the same positions, drop it to 0.16 to 0.33. A single statement that gives the queried key a rival value multiplies the loss by 9 to 110, and every additional rival adds to it. Inside the model, one quantity tracks the failure: the attention mass left on the evidence. Across 126,473 trials it explains up to 58% of the variance in loss; log length explains at most 8%. The masked-filler protocol behind the claim that length alone hurts reproduces on MMLU (11 to 18 points lost at 30K tokens), but fails a visible control in retrieval: masking costs 45 accuracy points at 32K tokens, while the same filler left visible costs none. In retrieval, context rot is a competition for attention. We release the probe and all 126K trial-level measurements.

Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter. We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks under five prompt variants each, measuring accuracy, token-probability calibration, verbal-confidence calibration, verbal parse rate, and prompt-perturbation spread for every (model x dataset x variant) cell. We find three broad results. First, evaluation design can materially change the conclusion. Switching Expected Calibration Error (ECE) token from a raw to a label-set-normalised definition changes per-cell calibration by a mean absolute 0.149. More strikingly, pairing a chain-of-thought prompt with a first-character evaluator on ARC-Challenge reduces apparent accuracy by 72-88% across all five primary models; two independent repair procedures recover 93.8% and 102.7% of the lost performance, indicating an evaluator-side rather than model-side failure. Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant. Third, prompt robustness does not track parameter count reliably. Across 10 instruct models, the correlation between model size and prompt-perturbation spread ranges from -0.244 to 0.474 across benchmarks. Taken together, these results show that reliability conclusions for small language models depend not only on the model being evaluated, but also on the evaluation pipeline used to measure it. We argue that calibration definitions, evaluator logic, verbal parseability, and prompt robustness should be reported explicitly when making reliability claims.

How can we train models to perform well on hard test data when hard training data is by definition difficult to label correctly? This question has been termed the scalable oversight problem and has drawn increasing attention as language models have continually improved. In this paper, we present the surprising conclusion that current language models often generalize relatively well from easy to hard data, even performing as well as "oracle" models trained on hard data. We demonstrate this kind of easy-to-hard generalization using simple training methods like in-context learning, linear classifier heads, and QLoRA for seven different measures of datapoint hardness, including six empirically diverse human hardness measures (like grade level) and one model-based measure (loss-based). Furthermore, we show that even if one cares most about model performance on hard data, it can be better to collect and train on easy data rather than hard data, since hard data is generally noisier and costlier to collect. Our experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. We conclude that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied, suggesting the scalable oversight problem may be easier than previously thought. Our code is available at https://github.com/allenai/easy-to-hard-generalization

With the growing importance of detecting misinformation, many studies have focused on verifying factual claims by retrieving evidence. However, canonical fact verification tasks do not apply to catching subtle differences in factually consistent claims, which might still bias the readers, especially on contentious political or economic issues. Our underlying assumption is that among the trusted sources, one's argument is not necessarily more true than the other, requiring comparison rather than verification. In this study, we propose ClaimDiff, a novel dataset that primarily focuses on comparing the nuance between claim pairs. In ClaimDiff, we provide 2,941 annotated claim pairs from 268 news articles. We observe that while humans are capable of detecting the nuances between claims, strong baselines struggle to detect them, showing over a 19% absolute gap with the humans. We hope this initial study could help readers to gain an unbiased grasp of contentious issues through machine-aided comparison.

The input side of the Information Bottleneck may contain task-irrelevant variation that still costs rate. We identify this cost exactly. Let T be a source and C a relevance variable. Suppose the deterministic statistic Z = phi(T) satisfies C-Z-T. For any encoder p(X|T), its conditional average over the fibres of phi preserves I(X;C) and lowers the rate by I(X;T|Z). The reverse pullback preserves both coordinates. These maps establish equality of the relevance-rate curves and Lagrangian infima on standard Borel spaces for every tradeoff parameter. They also characterise all attained optima. Every full-source minimiser factors through Z, and every reduced minimiser pulls back to T. When C is finite and distortion is logarithmic loss, replacement of T^n by Z^n also preserves the optimal remote distortion at every blocklength and message budget. The operational rate-distortion functions are therefore equal.

This study evaluates the performance of ChatGPT variants, GPT-3.5 and GPT-4, both with and without prompt engineering, against solely student work and a mixed category containing both student and GPT-4 contributions in university-level physics coding assignments using the Python language. Comparing 50 student submissions to 50 AI-generated submissions across different categories, and marked blindly by three independent markers, we amassed n = 300 data points. Students averaged 91.9% (SE:0.4), surpassing the highest performing AI submission category, GPT-4 with prompt engineering, which scored 81.1% (SE:0.8) - a statistically significant difference (p = 2.482 times 10^{-10}). Prompt engineering significantly improved scores for both GPT-4 (p = 1.661 times 10^{-4}) and GPT-3.5 (p = 4.967 times 10^{-9}). Additionally, the blinded markers were tasked with guessing the authorship of the submissions on a four-point Likert scale from `Definitely AI' to `Definitely Human'. They accurately identified the authorship, with 92.1% of the work categorized as 'Definitely Human' being human-authored. Simplifying this to a binary `AI' or `Human' categorization resulted in an average accuracy rate of 85.3%. These findings suggest that while AI-generated work closely approaches the quality of university students' work, it often remains detectable by human evaluators.

This paper investigates the fragility of Large Language Models (LLMs) in generalizing to novel inputs, specifically focusing on minor perturbations in well-established benchmarks (e.g., slight changes in question format or distractor length). Despite high benchmark scores, LLMs exhibit significant accuracy drops and unexpected biases (e.g., preference for longer distractors) when faced with these minor but content-preserving modifications. For example, Qwen 2.5 1.5B's MMLU score rises from 60 to 89 and drops from 89 to 36 when option lengths are changed without altering the question. Even GPT-4 experiences a 25-point accuracy loss when question types are changed, with a 6-point drop across all three modification categories. These analyses suggest that LLMs rely heavily on superficial cues rather than forming robust, abstract representations that generalize across formats, lexical variations, and irrelevant content shifts. This work aligns with the ACL 2025 theme track on the Generalization of NLP models, proposing a "Generalization Stress Test" to assess performance shifts under controlled perturbations. The study calls for reevaluating benchmarks and developing more reliable evaluation methodologies to capture LLM generalization abilities better.

Counterfactual prompting (i.e., perturbing a single factor and measuring output change) is widely used to evaluate things like LLM bias and CoT faithfulness. But in this work we argue that observed effects cannot be attributed to the targeted factor without accounting for baseline ``meaning-preserving'' modifications to text that establish general model sensitivity. This is because every counterfactual edit is a compound treatment that bundles the variable of interest with incidental surface-form variation; this violates treatment variation irrelevance. We observe prediction flip rates on MedQA of 14.9% when we surgically change patient gender. However, this is statistically indistinguishable from the flip rates induced by simply paraphrasing inputs (14.1%). In this case, it would therefore be unwarranted to conclude that the LLM is especially sensitive to patient gender. To account for this and robustly measure the effects of targeted interventions, we propose a framework in which we compare (via statistical testing) differences observed under target interventions to those induced by paraphrasing inputs. We then use this framework to revisit a analysis done on the MedPerturb dataset, which reported evidence of model sensitivity to patient demographics and stylistic cues. We find that these effects largely dissipate when we account for general model sensitivity, with only 5 of 120 tests reaching statistical significance. Applying the same framework to occupational biography classification, we detect clearly significant directional gender bias, showing that the framework identifies real directional effects even when they are small. We evaluate a range of metrics -- aggregate, per-sample distributional, and regression -- and find that per-sample metrics are dramatically more powerful than aggregate metrics and regression powerfully and uniquely characterizes effect direction and magnitude.

We describe the SemEval task of extracting keyphrases and relations between them from scientific documents, which is crucial for understanding which publications describe which processes, tasks and materials. Although this was a new task, we had a total of 26 submissions across 3 evaluation scenarios. We expect the task and the findings reported in this paper to be relevant for researchers working on understanding scientific content, as well as the broader knowledge base population and information extraction communities.

Null Hypothesis Significance Testing is the de facto tool for assessing effectiveness differences between Information Retrieval systems. Researchers use statistical tests to check whether those differences will generalise to online settings or are just due to the samples observed in the laboratory. Much work has been devoted to studying which test is the most reliable when comparing a pair of systems, but most of the IR real-world experiments involve more than two. In the multiple comparisons scenario, testing several systems simultaneously may inflate the errors committed by the tests. In this paper, we use a new approach to assess the reliability of multiple comparison procedures using simulated and real TREC data. Experiments show that Wilcoxon plus the Benjamini-Hochberg correction yields Type I error rates according to the significance level for typical sample sizes while being the best test in terms of statistical power.

We present E3, an automated review assistant that augments reviewers and engineering teams by identifying decision-relevant technical concerns in research papers. For each concern, E3 reports its nature, its location, its bearing on the contribution, and the analysis or evidence that would resolve it, covering unsupported claims, missing ablations, weak baselines, hidden assumptions, threats to validity, and leakage risks. To evaluate E3 without contamination confounds we adopt an issue-level backtesting protocol: the corpus is restricted to papers postdating the training cutoff of every automated source, and for each paper a meta-judge that observes only anonymised reviews labels every issue-source pair as Caught, Partial, or Missed. Applied to 100 ICLR 2026 papers and 4598 judged issue rows, comparing E3 against the ICLR human reviews and two prompt-matched LLM baselines built on gpt-5.4 from OpenAI and claude-opus-4-6 from Anthropic, with meta-judge gpt-5.5, E3 attains the highest recall on every aggregate metric. Partial-inclusive recall reaches 90.2 percent, which is 15.5 points over GPT, 17.1 points over Claude, and 29.2 points over the human reviews, and strict recall preserves the ordering at 65.8 percent. On concerns raised by the human reviewers, E3 recovers 89.6 percent; on concerns the human reviewers missed it surfaces 1635 additional rows admitted into the judged union, 406 above the next-best source. Corpus, baseline prompts, judge prompt template, and evaluation code are released.

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs' generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

The RVL-CDIP benchmark is widely used for measuring performance on the task of document classification. Despite its widespread use, we reveal several undesirable characteristics of the RVL-CDIP benchmark. These include (1) substantial amounts of label noise, which we estimate to be 8.1% (ranging between 1.6% to 16.9% per document category); (2) presence of many ambiguous or multi-label documents; (3) a large overlap between test and train splits, which can inflate model performance metrics; and (4) presence of sensitive personally-identifiable information like US Social Security numbers (SSNs). We argue that there is a risk in using RVL-CDIP for benchmarking document classifiers, as its limited scope, presence of errors (state-of-the-art models now achieve accuracy error rates that are within our estimated label error rate), and lack of diversity make it less than ideal for benchmarking. We further advocate for the creation of a new document classification benchmark, and provide recommendations for what characteristics such a resource should include.

We present ExAct, a new video-language benchmark for expert-level understanding of skilled physical human activities. Our new benchmark contains 3521 expert-curated video question-answer pairs spanning 11 physical activities in 6 domains: Sports, Bike Repair, Cooking, Health, Music, and Dance. ExAct requires the correct answer to be selected from five carefully designed candidate options, thus necessitating a nuanced, fine-grained, expert-level understanding of physical human skills. Evaluating the recent state-of-the-art VLMs on ExAct reveals a substantial performance gap relative to human expert performance. Specifically, the best-performing GPT-4o model achieves only 44.70% accuracy, well below the 82.02% attained by trained human specialists/experts. We believe that ExAct will be beneficial for developing and evaluating VLMs capable of precise understanding of human skills in various physical and procedural domains. Dataset and code are available at https://texaser.github.io/exact_project_page/

We evaluate language models' capability to generate novel research ideas through the lens of compression. We aim to lower-bound the potentially tiny probability that a language model generates the essence of a future research idea without any hints. Rather than estimate this probability through expensive repeated sampling, our Priced Guidance framework measures the compression cost: how many additional bits of information are needed to guide the model to recover the target idea. We prove that if the model can recover the target idea with at most K bits of guidance in expectation, then it can generate the idea without any guidance with probability at least 2^{-K}. In our framework, the language model, called the generator, can pose a sequence of multiple-choice questions and specify a probability distribution over possible answers. A guide, which is a language model with access to the target idea, selects answers. If a selected answer has prior probability p, the generator pays -log_2 p bits. The generator aims to produce an idea that matches the essence of the target idea with minimal cumulative cost. This cumulative cost equals, up to an additive constant, the number of bits of information sent by the guide. Using this methodology, we evaluate five generator LLMs (Opus 5, Fable 5.1, GPT-6 Astra, GPT-5.6 Sol, and GLM 5.3) on the core ideas in 87 recent high-quality deep learning papers and we use LLM as judge to determine whether the generated idea matches the target in terms of the central research object and defining mechanism. Fable 5.1 achieves the lowest median compression cost at 69.9 bits, substantially lower than gzip's median of 5,712 bits for losslessly compressing the summary of the target idea. A uniform ensemble of Fable 5.1, Opus 5, and Astra further reduces the median compression cost to 55.8 bits and improves the generation probability lower bound by 18,000 times.

Detailed profession-specific system prompts raise token use and estimated cost per response without a consistent accuracy gain. We evaluate Scientific Agents, an open-source corpus of 503 profession-specific AGENTS.md profiles, with Gemini 3.8 Flash via OpenRouter in the Pi agent harness. We compare matched profiles with four controls: a minimal baseline ("You are a helpful assistant"), the profile's opening role sentence, a generic scientific rigor guide, and a profile from an unrelated domain. Across nine text-based science benchmarks (4,531 sampled questions, 100 matched profiles), 4,488 items completed all five conditions after API-error retries, scored with automated, rule-based grading. The average profile-baseline accuracy difference is -0.6 percentage points (95% bootstrap interval [-1.5, +0.2] across fixed tasks), and no benchmark shows a statistically clear improvement. Matched profiles produced 1.5-2.3 times as many output tokens and cost 2.2-4.5 times more per successful call. On 60 tool-using BioMysteryBench bioinformatics problems (three runs each for baseline and profile), mean solve rates were 46.7% with the profile and 56.7% at baseline, a difference of -10.0 percentage points (95% interval [-16.7, -3.3]) driven by more frequent token- and time-limit stops under the profile. Longer prompts had one unexpected operational advantage: on SuperGPQA, frequent provider API drops left the short baseline with a correct first-pass answer on only 54.0% of items, against 71.6% with the profile. Generic and mismatched prompts were about as reliable, so this gain comes from prompt length or formatting rather than domain expertise. For the tested model and tasks, loading full profession profiles by default does not improve accuracy and costs considerably more; whether selective retrieval of profile sections or open-ended scientific tasks would change this remains to be tested.

In this paper, we present the main findings and compare the results of SemEval-2020 Task 10, Emphasis Selection for Written Text in Visual Media. The goal of this shared task is to design automatic methods for emphasis selection, i.e. choosing candidates for emphasis in textual content to enable automated design assistance in authoring. The main focus is on short text instances for social media, with a variety of examples, from social media posts to inspirational quotes. Participants were asked to model emphasis using plain text with no additional context from the user or other design considerations. SemEval-2020 Emphasis Selection shared task attracted 197 participants in the early phase and a total of 31 teams made submissions to this task. The highest-ranked submission achieved 0.823 Matchm score. The analysis of systems submitted to the task indicates that BERT and RoBERTa were the most common choice of pre-trained models used, and part of speech tag (POS) was the most useful feature. Full results can be found on the task's website.

Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model's 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

Since ChatGPT has emerged as a major AIGC model, providing high-quality responses across a wide range of applications (including software development and maintenance), it has attracted much interest from many individuals. ChatGPT has great promise, but there are serious problems that might arise from its misuse, especially in the realms of education and public safety. Several AIGC detectors are available, and they have all been tested on genuine text. However, more study is needed to see how effective they are for multi-domain ChatGPT material. This study aims to fill this need by creating a multi-domain dataset for testing the state-of-the-art APIs and tools for detecting artificially generated information used by universities and other research institutions. A large dataset consisting of articles, abstracts, stories, news, and product reviews was created for this study. The second step is to use the newly created dataset to put six tools through their paces. Six different artificial intelligence (AI) text identification systems, including "GPTkit," "GPTZero," "Originality," "Sapling," "Writer," and "Zylalab," have accuracy rates between 55.29 and 97.0%. Although all the tools fared well in the evaluations, originality was particularly effective across the board.

Recent advances in large language models (LLMs) have led to the development of various evaluation benchmarks. These benchmarks typically rely on a single instruction template for evaluating all LLMs on a specific task. In this paper, we comprehensively analyze the brittleness of results obtained via single-prompt evaluations across 6.5M instances, involving 20 different LLMs and 39 tasks from 3 benchmarks. To improve robustness of the analysis, we propose to evaluate LLMs with a set of diverse prompts instead. We discuss tailored evaluation metrics for specific use cases (e.g., LLM developers vs. developers interested in a specific downstream task), ensuring a more reliable and meaningful assessment of LLM capabilities. We then implement these criteria and conduct evaluations of multiple models, providing insights into the true strengths and limitations of current LLMs.

In text documents such as news articles, the content and key events usually revolve around a subset of all the entities mentioned in a document. These entities, often deemed as salient entities, provide useful cues of the aboutness of a document to a reader. Identifying the salience of entities was found helpful in several downstream applications such as search, ranking, and entity-centric summarization, among others. Prior work on salient entity detection mainly focused on machine learning models that require heavy feature engineering. We show that fine-tuning medium-sized language models with a cross-encoder style architecture yields substantial performance gains over feature engineering approaches. To this end, we conduct a comprehensive benchmarking of four publicly available datasets using models representative of the medium-sized pre-trained language model family. Additionally, we show that zero-shot prompting of instruction-tuned language models yields inferior results, indicating the task's uniqueness and complexity.

Scaling laws in language modeling traditionally quantify training loss as a function of dataset size and model parameters, providing compute-optimal estimates but often neglecting the impact of data quality on model generalization. In this paper, we extend the conventional understanding of scaling law by offering a microscopic view of data quality within the original formulation -- effective training tokens -- which we posit to be a critical determinant of performance for parameter-constrained language models. Specifically, we formulate the proposed term of effective training tokens to be a combination of two readily-computed indicators of text: (i) text diversity and (ii) syntheticity as measured by a teacher model. We pretrained over 200 models of 25M to 1.5B parameters on a diverse set of sampled, synthetic data, and estimated the constants that relate text quality, model size, training tokens, and eight reasoning task accuracy scores. We demonstrated the estimated constants yield +0.83 Pearson correlation with true accuracies, and analyzed it in scenarios involving widely-used data techniques such as data sampling and synthesis which aim to improve data quality.

Receiver Operating Characteristic (ROC) and Precision-Recall (PR) curves are fundamental tools for evaluating machine learning classifiers, offering detailed insights into the trade-offs between true positive rate vs. false positive rate (ROC) or precision vs. recall (PR). However, in Federated Learning (FL) scenarios, where data is distributed across multiple clients, computing these curves is challenging due to privacy and communication constraints. Specifically, the server cannot access raw prediction scores and class labels, which are used to compute the ROC and PR curves in a centralized setting. In this paper, we propose a novel method for approximating ROC and PR curves in a federated setting by estimating quantiles of the prediction score distribution under distributed differential privacy. We provide theoretical bounds on the Area Error (AE) between the true and estimated curves, demonstrating the trade-offs between approximation accuracy, privacy, and communication cost. Empirical results on real-world datasets demonstrate that our method achieves high approximation accuracy with minimal communication and strong privacy guarantees, making it practical for privacy-preserving model evaluation in federated systems.

Background: The positive predictive value (PPV) of large language model (LLM)-based proofreading for radiology reports is limited due to the low error prevalence. Purpose: To assess whether a three-pass LLM framework enhances PPV and reduces operational costs compared with baseline approaches. Materials and Methods: A retrospective analysis was performed on 1,000 consecutive radiology reports (250 each: radiography, ultrasonography, CT, MRI) from the MIMIC-III database. Two external datasets (CheXpert and Open-i) were validation sets. Three LLM frameworks were tested: (1) single-prompt detector; (2) extractor plus detector; and (3) extractor, detector, and false-positive verifier. Precision was measured by PPV and absolute true positive rate (aTPR). Efficiency was calculated from model inference charges and reviewer remuneration. Statistical significance was tested using cluster bootstrap, exact McNemar tests, and Holm-Bonferroni correction. Results: Framework PPV increased from 0.063 (95% CI, 0.036-0.101, Framework 1) to 0.079 (0.049-0.118, Framework 2), and significantly to 0.159 (0.090-0.252, Framework 3; P<.001 vs. baselines). aTPR remained stable (0.012-0.014; P>=.84). Operational costs per 1,000 reports dropped to USD 5.58 (Framework 3) from USD 9.72 (Framework 1) and USD 6.85 (Framework 2), reflecting reductions of 42.6% and 18.5%, respectively. Human-reviewed reports decreased from 192 to 88. External validation supported Framework 3's superior PPV (CheXpert 0.133, Open-i 0.105) and stable aTPR (0.007). Conclusion: A three-pass LLM framework significantly enhanced PPV and reduced operational costs, maintaining detection performance, providing an effective strategy for AI-assisted radiology report quality assurance.

Natural language conveys information at varying levels of granularity, from fine-grained references to broad descriptions. While granularity is fundamental to human communication, existing measures mostly capture surface detail or sentence specificity. We introduce Granuscore, a reference-free measure of granularity that leverages structural properties of a hierarchical embedding space. Granuscore reliably recovers hierarchical orderings on the Granola-EQ dataset and captures expected differences in granularity across discourse contexts. Across domains, we further show that Granuscore explains non-linear variation in sentence specificity beyond sentence length. Finally, we apply Granuscore to four question-answering benchmarks and analyze how granularity differs for questions, gold answers, and model outputs across response outcomes. The analysis reveals consistent differences in model behavior and provides a principled lens for characterizing the difficulty of QA datasets. Together, the results position Granuscore as a scalable, broadly applicable tool for analyzing granularity in text.

Behaviour-Driven Development (BDD) suites accumulate step-text duplication whose maintenance cost is established in prior work. Existing detection techniques require running the tests (Binamungu et al., 2018-2023) or are confined to a single organisation (Irshad et al., 2020-2022), leaving a gap: a purely static, paraphrase-robust, step-level detector usable on any repository. We fill the gap with cukereuse, an open-source Python CLI combining exact hashing, Levenshtein ratio, and sentence-transformer embeddings in a layered pipeline, released alongside an empirical corpus of 347 public GitHub repositories, 23,667 parsed .feature files, and 1,113,616 Gherkin steps. The step-weighted exact-duplicate rate is 80.2 %; the median-repository rate is 58.6 % (Spearman rho = 0.51 with size). The top hybrid cluster groups 20.7k occurrences across 2.2k files. Against 1,020 pairs manually labelled by the three authors under a released rubric (inter-annotator Fleiss' kappa = 0.84 on a 60-pair overlap), we report precision, recall, and F1 with bootstrap 95 % CIs under two protocols: the primary rubric and a score-free second-pass relabelling. The strongest honest pair-level number is near-exact at F1 = 0.822 on score-free labels; the primary-rubric semantic F1 = 0.906 is inflated by a stratification artefact that pins recall at 1.000. Lexical baselines (SourcererCC-style, NiCad-style) reach primary F1 = 0.761 and 0.799. The paper also presents a CDN-structured critique of Gherkin (Cognitive Dimensions of Notations); eight of fourteen dimensions are rated problematic or unsupported. The tool, corpus, labelled pairs, rubric, and pipeline are released under permissive licences.

Nearly all jurisdictions in the United States require a professional license exam, commonly referred to as "the Bar Exam," as a precondition for law practice. To even sit for the exam, most jurisdictions require that an applicant completes at least seven years of post-secondary education, including three years at an accredited law school. In addition, most test-takers also undergo weeks to months of further, exam-specific preparation. Despite this significant investment of time and capital, approximately one in five test-takers still score under the rate required to pass the exam on their first try. In the face of a complex task that requires such depth of knowledge, what, then, should we expect of the state of the art in "AI?" In this research, we document our experimental evaluation of the performance of OpenAI's `text-davinci-003` model, often-referred to as GPT-3.5, on the multistate multiple choice (MBE) section of the exam. While we find no benefit in fine-tuning over GPT-3.5's zero-shot performance at the scale of our training data, we do find that hyperparameter optimization and prompt engineering positively impacted GPT-3.5's zero-shot performance. For best prompt and parameters, GPT-3.5 achieves a headline correct rate of 50.3% on a complete NCBE MBE practice exam, significantly in excess of the 25% baseline guessing rate, and performs at a passing rate for both Evidence and Torts. GPT-3.5's ranking of responses is also highly-correlated with correctness; its top two and top three choices are correct 71% and 88% of the time, respectively, indicating very strong non-entailment performance. While our ability to interpret these results is limited by nascent scientific understanding of LLMs and the proprietary nature of GPT, we believe that these results strongly suggest that an LLM will pass the MBE component of the Bar Exam in the near future.

Scientific discovery is a cumulative process and requires new ideas to be situated within an ever-expanding landscape of existing knowledge. An emerging and critical challenge is how to identify conceptually relevant prior work from rapidly growing literature, and assess how a new idea differentiates from existing research. Current embedding approaches typically conflate distinct conceptual aspects into single representations and cannot support fine-grained literature retrieval; meanwhile, LLM-based evaluators are subject to sycophancy biases, failing to provide discriminative novelty assessment. To tackle these challenges, we introduce the Ideation Space, a structured representation that decomposes scientific knowledge into three distinct dimensions, i.e., research problem, methodology, and core findings, each learned through contrastive training. This framework enables principled measurement of conceptual distance between ideas, and modeling of ideation transitions that capture the logical connections within a proposed idea. Building upon this representation, we propose a Hierarchical Sub-Space Retrieval framework for efficient, targeted literature retrieval, and a Decomposed Novelty Assessment algorithm that identifies which aspects of an idea are novel. Extensive experiments demonstrate substantial improvements, where our approach achieves Recall@30 of 0.329 (16.7% over baselines), our ideation transition retrieval reaches Hit Rate@30 of 0.643, and novelty assessment attains 0.37 correlation with expert judgments. In summary, our work provides a promising paradigm for future research on accelerating and evaluating scientific discovery.

Recent advances in large language models and tool-using agents have expanded the range of benchmarked web tasks. Yet an important class of specialized retrieval tasks remains undercharacterized. On many specialized data-retrieval websites, answer-bearing evidence becomes accessible only after establishing the correct site-specific retrieval state through filters, views, hierarchies, or scopes. We term this capability state-gated retrieval (SGR). We introduce SGR-Bench, a benchmark for this setting containing 100 expert-curated tasks spanning six source families and 12 public data ecosystems. Each task requires discovering the appropriate website and configuring its site-specific retrieval state to produce a structured answer. SGR-Bench pairs constraint-guided and goal-oriented formulations of the same underlying problems, enabling controlled comparisons between explicit and implicit guidance for state-gated retrieval. We evaluate eight CLI-based agentic LLM systems and three commercial search-agent products. On SGR-Bench, the strongest system reaches only 66.18% item-level F1, while row-level F1 remains much lower. A manual audit of 156 analyzable failed CLI trajectories shows why: agents often reach a relevant web source, but establish the wrong site-specific retrieval state. Retrieval-scope drift (37.2%) and criterion mismatch (27.6%) dominate, whereas final answer composition accounts for only 10.3%. The dataset and single-case evaluation instructions are available at https://huggingface.co/datasets/PKUAIWeb/SGR-BENCH.

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d=28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.5% root mean square error (RMSE), (2) reconstruct the original total score with 0.8% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.93.

Despite the critical need to align search targets with users' intention, retrievers often only prioritize query information without delving into the users' intended search context. Enhancing the capability of retrievers to understand intentions and preferences of users, akin to language model instructions, has the potential to yield more aligned search targets. Prior studies restrict the application of instructions in information retrieval to a task description format, neglecting the broader context of diverse and evolving search scenarios. Furthermore, the prevailing benchmarks utilized for evaluation lack explicit tailoring to assess instruction-following ability, thereby hindering progress in this field. In response to these limitations, we propose a novel benchmark,INSTRUCTIR, specifically designed to evaluate instruction-following ability in information retrieval tasks. Our approach focuses on user-aligned instructions tailored to each query instance, reflecting the diverse characteristics inherent in real-world search scenarios. Through experimental analysis, we observe that retrievers fine-tuned to follow task-style instructions, such as INSTRUCTOR, can underperform compared to their non-instruction-tuned counterparts. This underscores potential overfitting issues inherent in constructing retrievers trained on existing instruction-aware retrieval datasets.

Keyword-matching benchmarks can credit small models for tool use they never perform. We document such a false positive in a matched-architecture pair of Spanish security language models and propose a ladder of strict, cheap diagnostics. A 661.6M parameter model (approx. 65% code/technical text; no dedicated SFT) and a 1,109M model (web-heavy multi-phase curriculum; 6B-token tool-SFT) share decoder, tokenizer, and special tokens, scoring almost identically on lenient tool-use metrics (B4: 0.660 vs. 0.650). Verbatim-reproduction checks on training examples separate them completely: the 600M emits valid tool calls with generalized arguments on 6/6 examples; the 1B does so on 0/6 across checkpoints. A first-token probe localizes the 1B's failure to a missing prior (prob. 10^{-4}--10^{-5} on <|tool_call|>), which was erased by its web-heavy training phase. A targeted SFT recipe (diverse corpus, 5x higher learning rate, 2,202 steps, ~3.3 GPU-hours) repairs the 1B using three orders of magnitude fewer tokens than the failed phase. On all 269 corpus rows, valid emission rises from 0.100 to 0.959 (600M: 0.926). On 238 unseen prompts, the repaired 1B passes 0.536 vs. the 600M's 0.428 (p = 0.004). Embedding-drift checks show the repair did not move the trigger token's tied embedding (97.7% of the bf16 table remains bit-identical), meaning changes live in the surrounding network. Both models over-trigger, rarely answering negative prompts without a call (0.09 for 600M, 0.17 for repaired 1B). Factorial analyses confirm all repair configurations install the format, though suppression benefits from a diverse corpus remain a hypothesis due to seed sensitivity. This cheap diagnostic ladder costs minutes of CPU time and should gate tool-use claims on small models.

To combat the misuse of Large Language Models (LLMs), many recent studies have presented LLM-generated-text detectors with promising performance. When users instruct LLMs to generate texts, the instruction can include different constraints depending on the user's need. However, most recent studies do not cover such diverse instruction patterns when creating datasets for LLM detection. In this paper, we reveal that even task-oriented constraints -- constraints that would naturally be included in an instruction and are not related to detection-evasion -- cause existing powerful detectors to have a large variance in detection performance. We focus on student essay writing as a realistic domain and manually create task-oriented constraints based on several factors for essay quality. Our experiments show that the standard deviation (SD) of current detector performance on texts generated by an instruction with such a constraint is significantly larger (up to an SD of 14.4 F1-score) than that by generating texts multiple times or paraphrasing the instruction. We also observe an overall trend where the constraints can make LLM detection more challenging than without them. Finally, our analysis indicates that the high instruction-following ability of LLMs fosters the large impact of such constraints on detection performance.

An agent that inherits six one-line memories may pull at most one archived source record before acting; a directive written into the store can steer that choice: a pointer to the record, a criterion that identifies it, or both. Across twelve registered studies on one instrument lineage (14,760 attempts) we measured where the request goes under each form. On six direct-provider models a length-matched criterion exceeded a bare id by +35.0 points [+31.2, +38.8] (Study D); the contrast failed its registered superiority rule on a nine-model OpenRouter-served panel (Study E). Appending the id cancelled the criterion on three Claude models (Opus 5: 40/40 to 0/40; Study F-x); six byte-matched edits gave each exact string its own effect (Study G), and a re-run at eighty runs per cell left fifteen of thirty replication contrasts within the margin, fifteen unresolved and none beyond (Study G'). A ratification line (+96.0 points on Opus 5) and a budget of two credits restored the target on all three (Study J); across five criterion strings the suffix's cancellation held for four of the five wordings on Opus 5 and all five wordings on Fable 5.1 (Study H2); in a second store every model followed the criterion (Study H1). Continued into a decision, the criterion moved the choice toward the current record (+100.0 points, Opus 5) and away from it on Fable 5.1 (Study I). A one-character plan pointer's effect (+78.0 points; Study B, after a correction of its first repository report) returned the same verdict under a prospectively registered re-run (+81.7 points; Study B'). All results are descriptive effects of exact edits on fixed panels with registered intervals and no mechanism claim.

The current trend to improve language model performance seems to be based on scaling up with the number of parameters (e.g. the state of the art GPT4 model has approximately 1.7 trillion parameters) or the amount of training data fed into the model. However this comes at significant costs in terms of computational resources and energy costs that compromise the sustainability of AI solutions, as well as risk relating to privacy and misuse. In this paper we present the Erasmian Language Model (ELM) a small context specific, 900 million parameter model, pre-trained and fine-tuned by and for Erasmus University Rotterdam. We show how the model performs adequately in a classroom context for essay writing, and how it achieves superior performance in subjects that are part of its context. This has implications for a wide range of institutions and organizations, showing that context specific language models may be a viable alternative for resource constrained, privacy sensitive use cases.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

Keyword extraction is the process of identifying the words or phrases that express the main concepts of text to the best of one's ability. Electronic infrastructure creates a considerable amount of text every day and at all times. This massive volume of documents makes it practically impossible for human resources to study and manage them. Nevertheless, the need for these documents to be accessed efficiently and effectively is evident in numerous purposes. A blog, news article, or technical note is considered a relatively long text since the reader aims to learn the subject based on keywords or topics. Our approach consists of a combination of two models: graph centrality features and textural features. The proposed method has been used to extract the best keyword among the candidate keywords with an optimal combination of graph centralities, such as degree, betweenness, eigenvector, closeness centrality and etc, and textural, such as Casing, Term position, Term frequency normalization, Term different sentence, Part Of Speech tagging. There have also been attempts to distinguish keywords from candidate phrases and consider them on separate keywords. For evaluating the proposed method, seven datasets were used: Semeval2010, SemEval2017, Inspec, fao30, Thesis100, pak2018, and Wikinews, with results reported as Precision, Recall, and F- measure. Our proposed method performed much better in terms of evaluation metrics in all reviewed datasets compared with available methods in literature. An approximate 16.9% increase was witnessed in F-score metric and this was much more for the Inspec in English datasets and WikiNews in forgone languages.

We introduce a high-quality dataset that contains 3,397 samples comprising (i) multiple choice questions, (ii) answers (including distractors), and (iii) their source documents, from the educational domain. Each question is phrased in two forms, normal and close. Correct answers are linked to source documents with sentence-level annotations. Thus, our versatile dataset can be used for both question and distractor generation, as well as to explore new challenges such as question format conversion. Furthermore, 903 questions are accompanied by their cognitive complexity level as per Bloom's taxonomy. All questions have been generated by educational experts rather than crowd workers to ensure they are maintaining educational and learning standards. Our analysis and experiments suggest distinguishable differences between our dataset and commonly used ones for question generation for educational purposes. We believe this new dataset can serve as a valuable resource for research and evaluation in the educational domain. The dataset and baselines will be released to support further research in question generation.

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

Neural networks -- especially those that use large, pre-trained language models -- have improved search engines in various ways. Most prominently, they can estimate the relevance of a passage or document to a user's query. In this work, we depart from this direction by exploring whether neural networks can effectively predict which of a document's passages are unlikely to be relevant to any query submitted to the search engine. We refer to this query-agnostic estimation of passage relevance as a passage's quality. We find that our novel methods for estimating passage quality allow passage corpora to be pruned considerably while maintaining statistically equivalent effectiveness; our best methods can consistently prune >25% of passages in a corpora, across various retrieval pipelines. Such substantial pruning reduces the operating costs of neural search engines in terms of computing resources, power usage, and carbon footprint -- both when processing queries (thanks to a smaller index size) and when indexing (lightweight models can prune low-quality passages prior to the costly dense or learned sparse encoding step). This work sets the stage for developing more advanced neural "learning-what-to-index" methods.

We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof"). The questions are also difficult for state-of-the-art AI systems, with our strongest GPT-4 based baseline achieving 39% accuracy. If we are to use future AI systems to help us answer very hard questions, for example, when developing new scientific knowledge, we need to develop scalable oversight methods that enable humans to supervise their outputs, which may be difficult even if the supervisors are themselves skilled and knowledgeable. The difficulty of GPQA both for skilled non-experts and frontier AI systems should enable realistic scalable oversight experiments, which we hope can help devise ways for human experts to reliably get truthful information from AI systems that surpass human capabilities.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

Selecting the ``right'' amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (https://huggingface.co/datasets/griffin/chain_of_density).

Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt elicitable if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer.

Much recent research on information retrieval has focused on how to transfer from one task (typically with abundant supervised data) to various other tasks where supervision is limited, with the implicit assumption that it is possible to generalize from one task to all the rest. However, this overlooks the fact that there are many diverse and unique retrieval tasks, each targeting different search intents, queries, and search domains. In this paper, we suggest to work on Few-shot Dense Retrieval, a setting where each task comes with a short description and a few examples. To amplify the power of a few examples, we propose Prompt-base Query Generation for Retriever (Promptagator), which leverages large language models (LLM) as a few-shot query generator, and creates task-specific retrievers based on the generated data. Powered by LLM's generalization ability, Promptagator makes it possible to create task-specific end-to-end retrievers solely based on a few examples {without} using Natural Questions or MS MARCO to train %question generators or dual encoders. Surprisingly, LLM prompting with no more than 8 examples allows dual encoders to outperform heavily engineered models trained on MS MARCO like ColBERT v2 by more than 1.2 nDCG on average on 11 retrieval sets. Further training standard-size re-rankers using the same generated data yields another 5.0 point nDCG improvement. Our studies determine that query generation can be far more effective than previously observed, especially when a small amount of task-specific knowledge is given.

System One models output calibrated probabilities over typed answers such as categorical choices, ordinal levels, or binary outcomes, via a non-generative interface. Software can act on these probabilities through thresholds, cost-weighted choices, and escalation rules. Consequently, if these probabilities are miscalibrated or wording-sensitive, the software ma take unintended actions leaving human operators with no textual rationale to inspect. Current evaluations largely report accuracy and calibration on public classification datasets without a clear reference. We present TypedBench, a benchmark for typed decision models like Jev built from seven policy-labelled generators and nine evaluation suites. We report accuracy as median and range across paraphrases, and calibration error relative to the finite-sample noise floor of a matched, perfectly calibrated predictor. We assess probability quality through selective prediction, ordinal proper scoring rules, and realised cost under asymmetric cost matrices. We evaluate a hosted model, an open encoder, and a family of open decoders spanning 0.8B-9B parameters on identical items. The hosted model follows the stated policy but is wording-sensitive and systematically underconfident; under asymmetric costs, using its probabilities can be worse than taking its top answer. Decoders route exactly and are slow as options or questions are added. The decoder is least accurate on policy questions and degrades with more options. Overall, typed decision models must be evaluated jointly on policy adherence, wording robustness, probability quality, and induced decision outcomes.

Literature search questions, such as "where can I find research on the evaluation of consistency in generated summaries?" pose significant challenges for modern search engines and retrieval systems. These questions often require a deep understanding of research concepts and the ability to reason over entire articles. In this work, we introduce LitSearch, a retrieval benchmark comprising 597 realistic literature search queries about recent ML and NLP papers. LitSearch is constructed using a combination of (1) questions generated by GPT-4 based on paragraphs containing inline citations from research papers and (2) questions about recently published papers, manually written by their authors. All LitSearch questions were manually examined or edited by experts to ensure high quality. We extensively benchmark state-of-the-art retrieval models and also evaluate two LLM-based reranking pipelines. We find a significant performance gap between BM25 and state-of-the-art dense retrievers, with a 24.8% difference in absolute recall@5. The LLM-based reranking strategies further improve the best-performing dense retriever by 4.4%. Additionally, commercial search engines and research tools like Google Search perform poorly on LitSearch, lagging behind the best dense retriever by 32 points. Taken together, these results show that LitSearch is an informative new testbed for retrieval systems while catering to a real-world use case.

If the same neural network architecture is trained multiple times on the same dataset, will it make similar linguistic generalizations across runs? To study this question, we fine-tuned 100 instances of BERT on the Multi-genre Natural Language Inference (MNLI) dataset and evaluated them on the HANS dataset, which evaluates syntactic generalization in natural language inference. On the MNLI development set, the behavior of all instances was remarkably consistent, with accuracy ranging between 83.6% and 84.8%. In stark contrast, the same models varied widely in their generalization performance. For example, on the simple case of subject-object swap (e.g., determining that "the doctor visited the lawyer" does not entail "the lawyer visited the doctor"), accuracy ranged from 0.00% to 66.2%. Such variation is likely due to the presence of many local minima that are equally attractive to a low-bias learner such as a neural network; decreasing the variability may therefore require models with stronger inductive biases.

Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 12 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 10 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information.

Large language models are increasingly used in scholarly work, yet it remains unclear whether their productivity gains are accompanied by changes in research novelty. We examine how relative abstract-level semantic novelty in Information Systems journals changed after ChatGPT became widely available and whether this change differed across institutional language contexts. We analyze 13,847 articles published from 2020 to 2025 in 44 A* and A Information Systems journals. Using SPECTER2 representations of titles and abstracts, we measure each article's semantic distance from its nearest recent predecessors and estimate a comparative pre/post model. Articles whose first authors were affiliated with institutions in non-English-dominant countries show a 0.176 standard deviation larger post-2022 decline in relative semantic novelty than articles from English-dominant affiliations, equivalent to about 7 percentile points. The pattern is similar across several alternative specifications, although the balanced-author estimate is less precise. We interpret this finding through a tension in generative AI-supported knowledge work. GenAI can widen access to prior knowledge and support new combinations, but it can also make established frames easier to reproduce. Because individual LLM use is not observed, the result identifies a heterogeneous post-2022 shift rather than an effect of LLM adoption. The study extends research on LLMs and scholarly productivity by shifting attention from publication counts to the semantic positioning of published articles and by showing that post-2022 change differs across institutional contexts.

Suppose Alice trains an open-weight language model and Bob uses a blackbox derivative of Alice's model to produce text. Can Alice prove that Bob is using her model, either by querying Bob's derivative model (query setting) or from the text alone (observational setting)? We formulate this question as an independence testing problem--in which the null hypothesis is that Bob's model or text is independent of Alice's randomized training run--and investigate it through the lens of palimpsestic memorization in language models: models are more likely to memorize data seen later in training, so we can test whether Bob is using Alice's model using test statistics that capture correlation between Bob's model or text and the ordering of training examples in Alice's training run. If Alice has randomly shuffled her training data, then any significant correlation amounts to exactly quantifiable statistical evidence against the null hypothesis, regardless of the composition of Alice's training data. In the query setting, we directly estimate (via prompting) the likelihood Bob's model gives to Alice's training examples and order; we correlate the likelihoods of over 40 fine-tunes of various Pythia and OLMo base models ranging from 1B to 12B parameters with the base model's training data order, achieving a p-value on the order of at most 1e-8 in all but six cases. In the observational setting, we try two approaches based on estimating 1) the likelihood of Bob's text overlapping with spans of Alice's training examples and 2) the likelihood of Bob's text with respect to different versions of Alice's model we obtain by repeating the last phase (e.g., 1%) of her training run on reshuffled data. The second approach can reliably distinguish Bob's text from as little as a few hundred tokens; the first does not involve any retraining but requires many more tokens (several hundred thousand) to achieve high power.

The validity screen (Cacioli, 2026d, 2026e) classifies LLM confidence signals as Valid, Indeterminate, or Invalid. We test whether these classifications predict selective prediction performance. Twenty frontier LLMs from seven families were evaluated on 524 items across six cognitive tracks. Valid models show mean Type 2 AUROC = .624 (SD = .048). Invalid models show mean AUROC = .357 (SD = .231). Cohen's d = 2.81, p = .002. The tiers order monotonically: Invalid (.357) < Indeterminate (.554) < Valid (.624). Split-half cross-validation yields median d = 1.77, P(d > 0) = 1.0 across 1,000 splits. The three-tier classification accounts for 47% of the variance in AUROC. DeepSeek-R1 drops from 85.3% accuracy at full coverage to 11.3% at 10% coverage. The screen predicts the criterion. For selective prediction, the screen matters.

Evolving documents, such as laws, tax codes, and software documentation, are amended, replaced, and sometimes reverted over time, so a question has different correct answers at different dates. In contrast to encyclopedic knowledge, where an old fact is simply overwritten, an amendment is itself an official text that states what it replaces and when it takes effect, and the earlier version stays correct for its validity period. The central challenge is therefore version resolution, that is, identifying the version in force on the queried date. Existing temporal QA datasets treat time only as an annotation, so version resolution stays untested. We present TIDE, an expert-verified benchmark of 3,050 QA pairs over 644 official customs instruments issued between 1969 and 2025 by the Government of Bangladesh, covering eight task types over deeply code-mixed documents that are heterogeneous in layout and dated in two calendars. In addition, we evaluate nine recent LLMs under a single protocol across parametric, gold-context, and retrieval access, scored by a three-judge LLM council with a hard date gate separating correct meaning from correct time. The best macro-averaged accuracy is only 68.5%. Resolving a version from an implicit date reaches 59.7%, and detecting that the supplied version does not govern the query reaches only 26.7%. Models are more likely to find correct versions than to reject incorrect ones, and they tend to follow a confident parametric answer over the supplied authoritative text. All code and data are available at https://github.com/icsetepa44/TIDE

Personalized summarization models cater to individuals' subjective understanding of saliency, as represented by their reading history and current topics of attention. Existing personalized text summarizers are primarily evaluated based on accuracy measures such as BLEU, ROUGE, and METEOR. However, a recent study argued that accuracy measures are inadequate for evaluating the degree of personalization of these models and proposed EGISES, the first metric to evaluate personalized text summaries. It was suggested that accuracy is a separate aspect and should be evaluated standalone. In this paper, we challenge the necessity of an accuracy leaderboard, suggesting that relying on accuracy-based aggregated results might lead to misleading conclusions. To support this, we delve deeper into EGISES, demonstrating both theoretically and empirically that it measures the degree of responsiveness, a necessary but not sufficient condition for degree-of-personalization. We subsequently propose PerSEval, a novel measure that satisfies the required sufficiency condition. Based on the benchmarking of ten SOTA summarization models on the PENS dataset, we empirically establish that -- (i) PerSEval is reliable w.r.t human-judgment correlation (Pearson's r = 0.73; Spearman's rho = 0.62; Kendall's tau = 0.42), (ii) PerSEval has high rank-stability, (iii) PerSEval as a rank-measure is not entailed by EGISES-based ranking, and (iv) PerSEval can be a standalone rank-measure without the need of any aggregated ranking.

To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%).

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.

Social highlighters let people mark passages that matter to them. We ask how much of an individual is recoverable from these naturalistic traces, using a co-readership identity control (the same document highlighted by many users) that holds document and topic fixed and asks whether a person's own history predicts their marks better than another reader's does. We separate generic salience (structure), crowd salience (what others marked), and personal salience (the individual residual). First, highlighting is social: which sentences you mark is predicted far better by the crowd than by structure or by a personal model, and even a well-estimated crowd, an information-privileged baseline that sees others' marks on the same document, beats a frontier LLM twin built from your other-document history; the within-document personal signal is at most a whisper (own-vs-other gap +0.017 by an embedding scorer, small but significant). Second, in sharp contrast, individuality lives in selection: asked which of the already-salient passages are yours, your own history is a strong, leakage-free predictor (gap +0.14). A topic decomposition shows this is largely stable thematic preference: it shrinks ~6-8x against a topically-matched peer, and a thin residual cannot be separated from finer topic. The non-obvious part is an asymmetry: under the same scorer the individual signal is ~6-8x weaker in salience than in selection. Methodologically, naive history-conditioning evaluations leak (the target's own marks enter the profile in ~42% of pairs, inflating personal scores by up to +0.15 AP) and small crowds overstate personalization; our results are leakage-free, use a dense crowd, and a model-matched control. Highlights carry a genuine individual signature, but a thin layer over a strong shared one, surfacing far more in which salient things a person selects than in what is salient.

We introduce Sabi\'a-2, a family of large language models trained on Portuguese texts. The models are evaluated on a diverse range of exams, including entry-level tests for Brazilian universities, professional certification exams, and graduate-level exams for various disciplines such as accounting, economics, engineering, law and medicine. Our results reveal that our best model so far, Sabi\'a-2 Medium, matches or surpasses GPT-4's performance in 23 out of 64 exams and outperforms GPT-3.5 in 58 out of 64 exams. Notably, specialization has a significant impact on a model's performance without the need to increase its size, allowing us to offer Sabi\'a-2 Medium at a price per token that is 10 times cheaper than GPT-4. Finally, we identified that math and coding are key abilities that need improvement.

Evaluation for automatic crisis reporting should measure what practitioners need, not what is convenient to compare. Through LLM-assisted, author-approved content analysis, we derive a 15-slot information schema from professional humanitarian situation reports using qualitative content analysis, then evaluate 121 human and 261 LLM-generated reports across five disasters and three hazard types. Under google/gemini-3.7-flash on Cyclone Idai, the best machine arm is nominally above the empirical human ceiling (0.827 vs. 0.821), but the difference is statistically indistinguishable from zero (95\% bootstrap CI [-.049,.059]; permutation p=.852), and the ordering does not replicate under openai/gpt-5.6-luna. On a blind 100-item Idai sample, the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's κ=.799. All four judges clear the plan's raw-agreement criterion of .7; only Gemini reaches .7 on chance-corrected κ. Across the three primary events, machine reports reproduce 5.4-11.0% of same-day official figures. Across those same events, raw exact-number matching links 14.3-19.6% of official figures and 85.4-90.1% of machine figures to the social stream; manual audit confirms 45/50 declared absences, 4/25 official matches, and 24/25 machine matches semantically. The issuing-provenance gap is positive in all 12 observed event-judge cells: report date, organisation, series, and authority were not supplied at professional-report coverage in any cell. On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.

We describe our system for LitTraceQA (GroundLM @ EMNLP 2026): given a research question, retrieve the relevant papers from a pool of 27,487, cite the page and the table or figure where the answer lives, and answer in a requested format. Our main finding is that evidence grounding is limited by retrieval, not by reading. The page selector put the annotator's page, which we call the gold page, in front of the model that locates evidence only about half the time (52.6% gold-page recall), while that model, given the page, cited the right one in 45 of the 48 locators it emitted (94%). When the page was missing it rarely said so: of 45 such cases it returned nothing 14 times, a wrong page 24 times, and a correct page 7 times, so the pipeline failed quietly almost twice as often as it failed visibly. Since the failure was that the right page was never shown, the fix is to stop choosing: each retrieved paper fits in the model's context, so we show it whole. Page ranking survives only as a fallback inside papers too long to fit, which no test-split paper was, and gold-page recall reaches 100% on the papers we can parse. Separately, questions that identify their target by position rather than content, such as "the first author of the 24th reference", are served by parsing rather than retrieval: we resolve the bibliography into an addressable list, which also supplies identifiers the evidence metric scores. The final system scores 0.762 paper F_1, 0.441 evidence F_1 and 0.920 multiple-choice accuracy on the held-out test split. Because the pipeline depends on a closed model without seed control, we release a harness that verifies the paper's central claims against committed artifacts.

A radiology report can already answer a clinical question, so it is hard to tell whether a vision-language model also uses the image. ModaLens, a paired image-swap audit, measures how report availability changes image sensitivity: MedGemma-27B on 3,199 paired MIMIC-CXR cases from 293 patients, all 14 questions per case (13 finding-specific and one composite), each image replaced by one from another study, usually of the same patient, with question and report fixed. Under an explicit answer instruction, the model's generated answer changes on 4.26 percent of trials with the report and 20.94 percent without it, a paired increase of 16.7 points (patient-clustered 95 percent CI 15.6 to 17.7), so report availability reduces image-swap sensitivity under this protocol; the original prompt with a lowercase first-token readout gives 4.70 percent against 17.07 percent, and substitutions also move continuous answer scores where the binary prediction does not change. The labels are derived from reports, which limits conclusions about visual correctness; the direction replicates in two further model lineages. Code, the exact prompts and a run record for every number are at https://github.com/criticaldata/MODALENS.

Sampling proper negatives from a large document pool is vital to effectively train a dense retrieval model. However, existing negative sampling strategies suffer from the uninformative or false negative problem. In this work, we empirically show that according to the measured relevance scores, the negatives ranked around the positives are generally more informative and less likely to be false negatives. Intuitively, these negatives are not too hard (may be false negatives) or too easy (uninformative). They are the ambiguous negatives and need more attention during training. Thus, we propose a simple ambiguous negatives sampling method, SimANS, which incorporates a new sampling probability distribution to sample more ambiguous negatives. Extensive experiments on four public and one industry datasets show the effectiveness of our approach. We made the code and models publicly available in https://github.com/microsoft/SimXNS.

Recent work finds that retrieval-augmented generation with large language models is prone to be influenced by the order of retrieved documents in the context. However, the lack of in-depth analysis limits the use of this phenomenon for prompt engineering in practice. In this study, we posit that likelihoods serve as an effective gauge for language model performance. Through experiments on two question-answering datasets with a variety of state-of-the-art language models, we reveal correlations between answer accuracy and the likelihood of the question at both the corpus level and the instance level. In addition, we find that question likelihood can also indicate the position of the task-relevant information in the context. Based on these findings, we propose two methods that use question likelihood as a gauge for selecting and constructing prompts that lead to better performance. We demonstrate their effectiveness with experiments. In addition, our likelihood-based methods are efficient, as they only need to compute the likelihood of the input, requiring much fewer language model passes than heuristic prompt engineering methods that require generating responses. Our analysis deepens our understanding of how input prompts affect model performance and provides a promising direction for efficient prompt optimization.

Practitioners make three prompt-design decisions with almost no controlled evidence behind them: how to format instructions and context (markdown, plain text, prose, or tabular), how many simultaneous instructions a system prompt can carry before compliance degrades, and how much context a model can hold before recall and honesty degrade. We report two controlled experiments crossing all three factors on one held, contamination-free synthetic corpus (the "Book of Veyra," 8,780 uniquely-named entities, deterministically regenerable from a fixed seed), evaluated across five models. Experiment 1 (960 calls/model) measures instruction-following decay as rule count N grows from 10 to 160, crossed with four formats and system-prompt vs. user-turn placement. Perfect-response rate collapses to zero by N=80 for every model, format, and placement. Placement produces effects at least as large as format at N=160 in most models, but the direction is model-specific. No model shows a reliable markdown advantage; one 35B model favors plain text instead. Experiment 2 (5,520 calls/model) measures recall accuracy, false-premise sycophancy, and absent-fact fabrication across a 2k-to-512k-token context ladder in the same four formats. Recall stays near ceiling through 64-128k tokens, then degrades sharply and format-dependently: one model's accuracy spread reaches 48 points at 128k tokens. Fabrication never occurs (0/5,760 probes), and sycophancy stays negligible (<=8.3%). What rises sharply near each model's context ceiling is outright refusal to answer (0% to 79-90%), distinct from sycophancy or fabrication. Neither pre-registered format ordering holds, and token overhead (+22% to +37% over plain text) further changes which format is preferable where accuracy spread is genuine. We release the full harness, corpus generator, and raw results (VeyraBench): https://github.com/iNetanel/veyrabench

The personal archive of Konstantin Tsiolkovsky (1857-1935) is held as fond 555 of the Archive of the Russian Academy of Sciences. The archive scanned the fond and published the images, but with no queryable catalogue, no full-text search and no dataset: the holdings can only be browsed one page at a time. This paper describes a machine-readable catalogue of all 2,019 files and 51,008 scans, a dating for 1,969 files taken from the archive's own descriptions, a page-level classification of every scan into handwriting and typescript, and a growing corpus of machine transcriptions (currently 322 files, 5,454 scans). It also reports a way to measure handwritten-text-recognition accuracy in an archive with no ground truth. Archives of the typewriter era often preserve one text twice, as manuscript and as a typed copy; transcribing both and comparing isolates the reading error, since source and pipeline are identical and only page difficulty differs. Across 294 such pairs from 27 files, two readings of a handwritten page agree on a median 37% of words. On two files that also have a published edition the estimate can be checked against ground truth: it is unbiased to within a percentage point and ranks pages as the truth does (rank correlation 0.92 where the edition is a faithful witness). This bounds use: two variants of one work here share 19% of words, below the rate at which two readings of a single page agree, so the redactions cannot be collated word by word at this quality. That negative result is reported as such, and the constraint is built into the tool.

Large language model (LLM) assistants are being deployed to clinicians across health systems, and judgments about their readiness rest largely on benchmark scores, most of them derived from examination questions or curated cases. A benchmark predicts performance in deployment only to the extent that its items resemble real use, yet whether benchmarks reflect the work these systems receive has rarely been measured. Here we analyze 127,833 queries sent by 6,342 physicians, advanced practice providers and nurses in 35 specialties to an institutional assistant during an eight-month roll-out. We characterize each query with RCQ-Map, a clinician-validated framework grounded in taxonomies of clinical questions and of LLM evaluation, which records its task, intent, answerability, missing information and potential harm. Documentation and administration (36.2%) and knowledge retrieval (28.9%) made up nearly two-thirds of use, and diagnosis 3.7%; more than a third of queries could not be answered well as posed. Applying RCQ-Map to 58 public benchmarks drawn from major evaluation suites and frontier model reports, which we assemble into the Clinical AI Benchmark Atlas, showed that the median benchmark contained no documentation requests and shared 31% of the task mix of real use, less than an even spread across task categories would. Benchmarks in suites designed to resemble clinical practice were individually no closer to real use than those used in frontier model reports. Benchmark scores therefore say little about how clinical AI performs on most of the work it is actually given, and evaluation should be matched to real clinical use.

Before deploying a language model (LM) within a given domain, it is important to measure its tendency to generate factually incorrect information in that domain. Existing factual generation evaluation methods focus on facts sampled from the LM itself, and thus do not control the set of evaluated facts and might under-represent rare and unlikely facts. We propose FACTOR: Factual Assessment via Corpus TransfORmation, a scalable approach for evaluating LM factuality. FACTOR automatically transforms a factual corpus of interest into a benchmark evaluating an LM's propensity to generate true facts from the corpus vs. similar but incorrect statements. We use our framework to create two benchmarks: Wiki-FACTOR and News-FACTOR. We show that: (i) our benchmark scores increase with model size and improve when the LM is augmented with retrieval; (ii) benchmark score correlates with perplexity, but the two metrics do not always agree on model ranking; and (iii) when perplexity and benchmark score disagree, the latter better reflects factuality in open-ended generation, as measured by human annotators. We make our data and code publicly available in https://github.com/AI21Labs/factor.

Machine learning methods have proven useful in transcribing historical data. However, results from even highly accurate methods require manual verification and correction. Such manual review can be time-consuming and expensive, therefore the objective of this paper was to make it more efficient. Previously, we used machine learning to transcribe 2.3 million handwritten occupation codes from the Norwegian 1950 census with high accuracy (97%). We manually reviewed the 90,000 (3%) codes with the lowest model confidence. We allocated those 90,000 codes to human reviewers, who used our annotation tool to review the codes. To assess reviewer agreement, some codes were assigned to multiple reviewers. We then analyzed the review results to understand the relationship between accuracy improvements and effort. Additionally, we interviewed the reviewers to improve the workflow. The reviewers corrected 62.8% of the labels and agreed with the model label in 31.9% of cases. About 0.2% of the images could not be assigned a label, while for 5.1% the reviewers were uncertain, or they assigned an invalid label. 9,000 images were independently reviewed by multiple reviewers, resulting in an agreement of 86.43% and disagreement of 8.96%. We learned that our automatic transcription is biased towards the most frequent codes, with a higher degree of misclassification for the lowest frequency codes. Our interview findings show that the reviewers did internal quality control and found our custom tool well-suited. So, only one reviewer is needed, but they should report uncertainty.