Daily Papers

Multiple-choice question answering (MCQA) is often used to evaluate large language models (LLMs). To see if MCQA assesses LLMs as intended, we probe if LLMs can perform MCQA with choices-only prompts, where models must select the correct answer only from the choices. In three MCQA datasets and four LLMs, this prompt bests a majority baseline in 11/12 cases, with up to 0.33 accuracy gain. To help explain this behavior, we conduct an in-depth, black-box analysis on memorization, choice dynamics, and question inference. Our key findings are threefold. First, we find no evidence that the choices-only accuracy stems from memorization alone. Second, priors over individual choices do not fully explain choices-only accuracy, hinting that LLMs use the group dynamics of choices. Third, LLMs have some ability to infer a relevant question from choices, and surprisingly can sometimes even match the original question. We hope to motivate the use of stronger baselines in MCQA benchmarks, the design of robust MCQA datasets, and further efforts to explain LLM decision-making.

In this paper we study yes/no questions that are naturally occurring --- meaning that they are generated in unprompted and unconstrained settings. We build a reading comprehension dataset, BoolQ, of such questions, and show that they are unexpectedly challenging. They often query for complex, non-factoid information, and require difficult entailment-like inference to solve. We also explore the effectiveness of a range of transfer learning baselines. We find that transferring from entailment data is more effective than transferring from paraphrase or extractive QA data, and that it, surprisingly, continues to be very beneficial even when starting from massive pre-trained language models such as BERT. Our best method trains BERT on MultiNLI and then re-trains it on our train set. It achieves 80.4% accuracy compared to 90% accuracy of human annotators (and 62% majority-baseline), leaving a significant gap for future work.

Recent research in disaster informatics demonstrates a practical and important use case of artificial intelligence to save human lives and suffering during natural disasters based on social media contents (text and images). While notable progress has been made using texts, research on exploiting the images remains relatively under-explored. To advance image-based approaches, we propose MEDIC (Available at: https://crisisnlp.qcri.org/medic/index.html), which is the largest social media image classification dataset for humanitarian response consisting of 71,198 images to address four different tasks in a multi-task learning setup. This is the first dataset of its kind: social media images, disaster response, and multi-task learning research. An important property of this dataset is its high potential to facilitate research on multi-task learning, which recently receives much interest from the machine learning community and has shown remarkable results in terms of memory, inference speed, performance, and generalization capability. Therefore, the proposed dataset is an important resource for advancing image-based disaster management and multi-task machine learning research. We experiment with different deep learning architectures and report promising results, which are above the majority baselines for all tasks. Along with the dataset, we also release all relevant scripts (https://github.com/firojalam/medic).

We introduce PubMedQA, a novel biomedical question answering (QA) dataset collected from PubMed abstracts. The task of PubMedQA is to answer research questions with yes/no/maybe (e.g.: Do preoperative statins reduce atrial fibrillation after coronary artery bypass grafting?) using the corresponding abstracts. PubMedQA has 1k expert-annotated, 61.2k unlabeled and 211.3k artificially generated QA instances. Each PubMedQA instance is composed of (1) a question which is either an existing research article title or derived from one, (2) a context which is the corresponding abstract without its conclusion, (3) a long answer, which is the conclusion of the abstract and, presumably, answers the research question, and (4) a yes/no/maybe answer which summarizes the conclusion. PubMedQA is the first QA dataset where reasoning over biomedical research texts, especially their quantitative contents, is required to answer the questions. Our best performing model, multi-phase fine-tuning of BioBERT with long answer bag-of-word statistics as additional supervision, achieves 68.1% accuracy, compared to single human performance of 78.0% accuracy and majority-baseline of 55.2% accuracy, leaving much room for improvement. PubMedQA is publicly available at https://pubmedqa.github.io.

AI systems for peer review fail on three fronts: they train on Computer Science and Machine Learning venues alone, ignore the iterative dialogue that validates science, and evaluate on stylistic mimicry rather than real editorial judgment. We introduce FirstPass, a dataset and fine-tuned model that addresses all three. Curating 3,668 complete multi-round peer-review dialogues from Nature Communications across five scientific domains (biology, chemistry, neuroscience, physics, and earth science), we exploit mandatory transparent peer review (instituted November 2022) and verify 100% content integrity by automated audit. We fine-tune Qwen2.5-7B-Instruct via Low-Rank Adaptation (LoRA) on three tasks: review generation, reviewer updating, and revision-cycle prediction. Our key finding is that response-only loss masking is a prerequisite, not an optimization: without it, accuracy is 62.0%, below the majority baseline; with it, FirstPass achieves 80.5% accuracy and F1-macro 78.2% on predicting editorial outcomes (Standard vs. Extended revision cycles), outperforming Gemini-3.1-flash-lite-preview zero-shot by 10.4 percentage points and all baselines with statistical significance (McNemar p < 0.001). On generation, FirstPass produces reviews averaging 1,187 words, substantially closer to human references (2,155 words) than any baseline, achieving ROUGE-L 0.154 with significant gains over Qwen and DeepSeek zero-shot (p < 0.001). Deployed in the pre-submission loop as an anticipatory scientific co-author, FirstPass simulates expert critique and predicts revision cycle outcomes before submission, giving authors the judgment a trusted colleague would provide, with consistent cross-domain performance across five disciplines.

Peer reviewing is a central component in the scientific publishing process. We present the first public dataset of scientific peer reviews available for research purposes (PeerRead v1) providing an opportunity to study this important artifact. The dataset consists of 14.7K paper drafts and the corresponding accept/reject decisions in top-tier venues including ACL, NIPS and ICLR. The dataset also includes 10.7K textual peer reviews written by experts for a subset of the papers. We describe the data collection process and report interesting observed phenomena in the peer reviews. We also propose two novel NLP tasks based on this dataset and provide simple baseline models. In the first task, we show that simple models can predict whether a paper is accepted with up to 21% error reduction compared to the majority baseline. In the second task, we predict the numerical scores of review aspects and show that simple models can outperform the mean baseline for aspects with high variance such as 'originality' and 'impact'.

Large language models (LLMs) offer impressive performance in various zero-shot and few-shot tasks. However, their success in zero-shot and few-shot settings may be affected by task contamination, a potential limitation that has not been thoroughly examined. This paper investigates how zero-shot and few-shot performance of LLMs has changed chronologically over time. Utilizing GPT-3 series models and several other recent open-sourced LLMs, and controlling for dataset difficulty, we find that on datasets released before the LLM training data creation date, LLMs perform surprisingly better than on datasets released after. This strongly indicates that, for many LLMs, there exists task contamination on zero-shot and few-shot evaluation for datasets released prior to the LLMs' training data creation date. Additionally, we utilize training data inspection, task example extraction, and a membership inference attack, which reveal further evidence of task contamination. Importantly, we find that for classification tasks with no possibility of task contamination, LLMs rarely demonstrate statistically significant improvements over simple majority baselines, in both zero and few-shot settings.

Aligned language models gate behaviors such as refusal and language routing through sparse feed forward neurons, yet no theory predicts when a single neuron intervention controls a behavior coherently rather than collapsing the output. We develop a budget normalized control window framework for single neuron steering. A dose along one write direction reduces to one control coordinate: the alignment between the residual stream and the write, driven along a universal saturation curve in units of a coherence budget set by the residual norm divided by the write norm. Coherent control exists when a behavior trigger lies below the collapse ceiling. The same coordinate governs benign mode switches and refusal; the ceiling follows from weights and one generic forward pass, while triggers are measured at rollout. On fifteen held out neurons, the predicted ceiling has mean absolute error 0.14, about 0.07 in bulk layers, and the committed open or closed verdict holds on eleven against a ten of fifteen majority baseline. Closed cases expose three failure modes rather than violations: collapse before trigger, too little depth to propagate, or a normalization that caps how far one neuron can push. The law explains why local gradient attribution anti predicts control: true controllers write off the readout axis and carry a near zero first order gradient. A forward only contrastive screen made precise by the window recovers controllers that attribution misses. On refusal, the hardest case, intervention success is typed, not scalar: coherent bypass and strict actionable reach separate, so a neuron can flip refusal in fluent, on task text with no actionable content, and genuine actionable reach appears only for three of six audited Llama pivots and only at later rollout horizons. Single neuron steering is therefore a budgeted, typed audit of controllability rather than a fixed dose anecdote.

Vision-Language-Action models (VLAs) enable general-purpose embodied intelligence, but their performance is brittle under distribution shift, where unanticipated failures can damage equipment or endanger people. Deploying VLAs in real-world systems therefore requires a runtime monitor that is timely, calibrated, and informative while being efficient enough to run on the edge. Existing methods fail to fulfill these requirements. We present VLA-Tug, a lightweight probe over the latent layers of a VLA backbone that improves early failure detection and bounds the missed-detection and false-alarm rates, with minimal inference overhead. We show that supervising on the pre-failure segment of a trajectory increases early failure detection at matched false-positive rates by up to 30% over the strongest learned baseline. Under shift to unseen tasks, the bounds on missed-detection and false-alarm rates remain valid in up to 86% of runs, versus at most 21% for baselines. Unlike existing methods, VLA-Tug predicts the failure mode to enable corrective actions, raising top-3 accuracy on unseen tasks by up to 44% over a majority baseline. VLA-Tug enables real-time monitoring where the well-timed and informative probe supports lightweight steering that improves task success. We release a dataset of 3,180 annotated manipulation episodes to support future work on latent failure detection in VLAs at https://vla-tug.github.io.

We propose a hypothesis only baseline for diagnosing Natural Language Inference (NLI). Especially when an NLI dataset assumes inference is occurring based purely on the relationship between a context and a hypothesis, it follows that assessing entailment relations while ignoring the provided context is a degenerate solution. Yet, through experiments on ten distinct NLI datasets, we find that this approach, which we refer to as a hypothesis-only model, is able to significantly outperform a majority class baseline across a number of NLI datasets. Our analysis suggests that statistical irregularities may allow a model to perform NLI in some datasets beyond what should be achievable without access to the context.

While modern autonomous driving systems excel at perception tasks such as object detection and trajectory prediction, they lack the high-level causal reasoning required to interpret traffic accidents. In particular, determining responsibility, such as identifying who is at fault and which traffic rule was violated, remains largely unexplored in current benchmarks. To this end, we introduce CAViAR (Causal Accident Video and Incident Analysis Repository), a human-annotated dashcam benchmark comprising 2,249 real-world accident videos collected from CarCrashDataset (CCD) and Nexar. Each video is annotated with structured labels spanning environmental conditions, accident type, causal explanation, apparent At-Fault Agent, affected agent, and apparent rule-violation category. We benchmark state-of-the-art vision-language models (VLMs), including Cosmos-Reason2, Qwen3-VL, and InternVL3. Once class imbalance is accounted for with majority/random baselines and balanced metrics, perceptual competence is uneven--lighting is nearly solved, whereas weather and road-condition accuracy fall at or below the majority-class baseline---and all models degrade sharply on accident type and responsibility reasoning. Overall, CAViAR exposes a practical Perception--Reasoning Gap: current VLMs may recognize salient context, but do not reliably map visible agent actions to annotated rule-relevant responsibility categories in safety-critical driving scenarios. Code, annotation schema, prompts, and evaluation scripts are available at: https://github.com/nec-labs-ma/CAViAR

A robot that fails at a task faces the first decision in corrective dialogue: act on its own diagnosis, consult another onboard sensor, or interrupt a person. Choosing well requires knowing how much the robot's sensors reveal about the cause and how reliable the robot's own diagnosis is. We build a simulated benchmark in which every failure's true cause is known, because we injected it, and measure what each sensor reveals, with explicit checks against data leakage. Some failures are diagnosable from camera images; others only from the robot's force data (0.99 from force data, no image method above 0.55). We then test six open vision-language models. Their behavior tracks the surface of the prompt, not the evidence: moving the refusal option from last to first in the answer list collapses refusal rates from 78-100% to 0-6% in three of the six swept model-and-family pairs. Accuracy from frames stays at or below a majority-class baseline under every prompt variant, with or without worked examples, and stated confidence carries no information about correctness. Handing the same models the force data as ten lines of text produces the first above-baseline diagnoses, in four of the six models: much of the failure reflects missing sensor data, not missing ability. We pose the choice as a three-action decision problem, act, consult your own sensors, or ask a human, whose optimal policy follows from measured accuracy. The models do not follow it, and their ask rates ignore a fourfold change in question cost. One question to a human still lifts them from that baseline to roughly the answerer's own reliability (0.70-0.81 when they ask). The decision to ask should be tied to measured accuracy and stated costs, not to the model's confidence.

India's Right to Information Act, 2005 gives every citizen the right to demand information from public authorities, yet in practice most people cannot make sense of the dense administrative language used in Central Information Commission (CIC) decisions, let alone predict whether an appeal is worth filing. This paper introduces RTI-Bench, a structured dataset of CIC decisions with outcome labels, exemption citations, IRAC-style reasoning components, and procedural timelines. To the best of our knowledge it is the first publicly released structured dataset for Indian RTI administrative decisions. The dataset draws from two sources: 1,218 cases from a publicly available instruction-response corpus (with structured fields added through rule-based extraction), and 298 CIC decision PDFs collected directly from the Commission portal, spanning five commissioners and three document format generations from 2023 to 2026. Label coverage reaches 89% on the instruction-response corpus. For the PDF subset of 239 primary decisions, coverage is 51% in this first release. A random sample of 50 labelled cases was manually reviewed, yielding a label precision of 95.3%. A zero-shot Mistral 7B baseline on 100 cases gives 57.3% accuracy and 37.0% macro-F1 on outcome prediction, well above the majority-class baseline of 14.3% macro-F1. RTI-Bench is available at https://huggingface.co/datasets/joyboseroy/rti-bench

This paper presents our winning submission to the AI Mathematical Olympiad - Progress Prize 2 (AIMO-2) competition. Our recipe for building state-of-the-art mathematical reasoning models relies on three key pillars. First, we create a large-scale dataset comprising 540K unique high-quality math problems, including olympiad-level problems, and their 3.2M long-reasoning solutions. Second, we develop a novel method to integrate code execution with long reasoning models through iterative training, generation, and quality filtering, resulting in 1.7M high-quality Tool-Integrated Reasoning solutions. Third, we create a pipeline to train models to select the most promising solution from many candidates. We show that such generative solution selection (GenSelect) can significantly improve upon majority voting baseline. Combining these ideas, we train a series of models that achieve state-of-the-art results on mathematical reasoning benchmarks. To facilitate further research, we release our code, models, and the complete OpenMathReasoning dataset under a commercially permissive license.

This paper describes the results of SemEval 2023 task 7 -- Multi-Evidence Natural Language Inference for Clinical Trial Data (NLI4CT) -- consisting of 2 tasks, a Natural Language Inference (NLI) task, and an evidence selection task on clinical trial data. The proposed challenges require multi-hop biomedical and numerical reasoning, which are of significant importance to the development of systems capable of large-scale interpretation and retrieval of medical evidence, to provide personalized evidence-based care. Task 1, the entailment task, received 643 submissions from 40 participants, and Task 2, the evidence selection task, received 364 submissions from 23 participants. The tasks are challenging, with the majority of submitted systems failing to significantly outperform the majority class baseline on the entailment task, and we observe significantly better performance on the evidence selection task than on the entailment task. Increasing the number of model parameters leads to a direct increase in performance, far more significant than the effect of biomedical pre-training. Future works could explore the limitations of large models for generalization and numerical inference, and investigate methods to augment clinical datasets to allow for more rigorous testing and to facilitate fine-tuning. We envisage that the dataset, models, and results of this task will be useful to the biomedical NLI and evidence retrieval communities. The dataset, competition leaderboard, and website are publicly available.

LLMs are routinely evaluated on language use, yet their explicit knowledge about linguistic structure remains poorly understood. Existing linguistic benchmarks focus on narrow phenomena, emphasize high-resource languages, and rarely test metalinguistic knowledge - explicit reasoning about language structure. We present a multilingual evaluation of metalinguistic knowledge in LLMs, based on the World Atlas of Language Structures (WALS), documenting 192 linguistic features across 2,660 languages. We convert WALS features into natural-language multiple-choice questions and evaluate models across documented languages. Using accuracy and macro F1, and comparing to chance and majority-class baselines, we assess performance and analyse variation across linguistic domains and language-related factors. Results show limited metalinguistic knowledge: GPT-4o performs best but achieves moderate accuracy (0.367), while open-source models lag. Although all models perform above chance, they fail to outperform the majority-class baseline, suggesting they capture broad cross-linguistic patterns but lack fine-grained distinctions. Performance varies by domain, partly reflecting differences in online visibility. At the language level, accuracy correlates with digital language status: languages with greater digital presence and resources are evaluated more accurately, while low-resource languages perform worse. Analysis of predictive factors confirms that resource-related indicators (Wikipedia size, corpus availability) are more informative than geographic, genealogical, or sociolinguistic factors. Overall, LLM metalinguistic knowledge appears fragmented and shaped mainly by data availability, rather than broadly generalizable grammatical competence. We release the benchmark as an open-source dataset to support evaluation across languages and encourage greater global linguistic diversity in future LLMs.

Network operators frequently encounter protocol divergence---where ICMP, TCP, and UDP experience different treatment along a path. Existing tools identify where divergence occurs but not why, forcing manual investigation that delays resolution. We present two contributions: (1) the AS-Boundary Divergence Score (ABDS), a metric that correlates protocol divergence locations with BGP topology to automatically classify causes as routing-policy-driven versus internal-policy-driven; and (2) the localhost response signature, an empirical finding that security appliances return 127.0.0.1 in TTL-exceeded messages to mask their IPs. We evaluate ABDS against 496 targets across 8 categories (Finance, Government, Education, Healthcare, Tech, E-commerce, Media, International), with multi-vantage validation from 10+ global locations confirming routing consistency. ABDS achieves 100% classification accuracy on verified divergent targets (50/50), significantly outperforming majority-class baseline (54%, McNemar's p < 0.001). The localhost signature achieves 100% precision (18/18, 95% CI: [82%, 100%]) in identifying security appliances across 9 categories including government (IRS, USPS, VA), finance, and healthcare. Our open-source Rust implementation enables automated escalation decisions, reducing mean-time-to-resolution for network operators.

Omni models can describe video content, but can they locate events in time, preserve event order, and judge audio-visual synchronization? We introduce AVTrace (Audio-Visual Temporal Reasoning Assessment and Capability Evaluation), a silver-standard diagnostic suite spanning onset and span grounding, synchronization, next-step prediction, cross-modal localization, chain parsing, and event-conditioned comprehension. It contains 34,114 training examples and category-balanced development and test splits of 3,500 and 7,000 examples. We evaluate five open omni models under their respective input configurations using reference-blind response normalization followed by deterministic scoring. All five off-the-shelf systems score below the test split's majority-label baseline of 0.556 on synchronization verification, and obtain low scores on chain parsing and event-conditioned grounding and comprehension. Development-set perturbations reveal task-dependent sensitivity in Qwen3-Omni-30B to modality removal and changes in visual input processing, without isolating their underlying causes. Parameter-efficient temporal post-training improves Gemma4-E4B-it on several benchmark metrics. On three external image benchmarks, task metrics change modestly, including some degradations, while teacher-forcing perplexity decreases. Together, these findings show that semantic reference-text overlap should not be treated as a proxy for temporal localization, and that AVTrace can identify task-specific weaknesses while providing a testbed for temporal post-training.

Process Reward Models (PRMs) have proven effective at enhancing mathematical reasoning for Large Language Models (LLMs) by leveraging increased inference-time computation. However, they are predominantly trained on mathematical data and their generalizability to non-mathematical domains has not been rigorously studied. In response, this work first shows that current PRMs have poor performance in other domains. To address this limitation, we introduce VersaPRM, a multi-domain PRM trained on synthetic reasoning data generated using our novel data generation and annotation method. VersaPRM achieves consistent performance gains across diverse domains. For instance, in the MMLU-Pro category of Law, VersaPRM via weighted majority voting, achieves a 7.9% performance gain over the majority voting baseline -- surpassing Qwen2.5-Math-PRM's gain of 1.3%. We further contribute to the community by open-sourcing all data, code and models for VersaPRM.

Quantitative reasoning is a higher-order reasoning skill that any intelligent natural language understanding system can reasonably be expected to handle. We present EQUATE (Evaluating Quantitative Understanding Aptitude in Textual Entailment), a new framework for quantitative reasoning in textual entailment. We benchmark the performance of 9 published NLI models on EQUATE, and find that on average, state-of-the-art methods do not achieve an absolute improvement over a majority-class baseline, suggesting that they do not implicitly learn to reason with quantities. We establish a new baseline Q-REAS that manipulates quantities symbolically. In comparison to the best performing NLI model, it achieves success on numerical reasoning tests (+24.2%), but has limited verbal reasoning capabilities (-8.1%). We hope our evaluation framework will support the development of models of quantitative reasoning in language understanding.

Machine learning classifiers for bearing fault detection produce scalar confidence scores that conflate confident errors with genuinely ambiguous predictions, and the conventional truth/falsity pair (F = 1 - T) is algebraically redundant by construction. We operationalize a refined neutrosophic decomposition of a Random Forest + XGBoost + Logistic Regression ensemble into four indicators -- T-hat (top-class evidence), F-hat (best-competitor evidence), predictive entropy I1-hat, and decision disagreement I2-hat -- evaluated on two bearing benchmarks (CWRU and JNU, 600-1000 rpm) under a leave-one-condition-out protocol. On CWRU, after correcting a file-to-class mapping error, the ensemble reaches 100.00 percent accuracy on three of four held-out loads (92.27 percent on the fourth), leaving too few errors for uncertainty analysis. On JNU, holding out 1000 rpm, accuracy collapses to 40.64 percent, below a majority-class baseline; Logistic Regression (57.91 percent) generalizes far better than the tree ensembles. I1-hat shows a robust association with error beyond T-hat/F-hat, while I2-hat contributes little; standalone Logistic Regression confidence outperforms the full decomposition, a boundary condition we report honestly. Two further results extend this: fusing a time-domain and a frequency-domain model of the same signal and scoring their Jensen-Shannon divergence beats that model own entropy (AURC 0.29 vs. 0.36 on the standard split; 0.54 vs. 0.73 under a harder single-condition reproduction), the only indicator moving correctly under a CWRU-versus-JNU distributional-shift contrast; and, on CWRU alone, literature-verified bearing fault frequencies, correctly demodulated via the envelope spectrum, separate most fault classes almost perfectly (99.57 percent) using three interpretable features. Code, logs, and figures are released for independent verification.

Process Reward Models (PRMs) are a powerful mechanism for steering large language model reasoning by providing fine-grained, step-level supervision. However, this effectiveness comes at a significant cost: PRMs require expert annotations for every reasoning step, making them costly and difficult to scale. Here, we propose a method for training unsupervised PRMs (uPRM) that requires no human supervision, neither at the level of step-by-step annotations nor through ground-truth verification of final answers. The key idea behind our approach is to define a scoring function, derived from LLM next-token probabilities, that jointly assesses candidate positions of first erroneous steps across a batch of reasoning trajectories. We demonstrate the effectiveness of uPRM across diverse scenarios: (i) uPRM achieves up to 15% absolute accuracy improvements over the LLM-as-a-Judge in identifying first erroneous steps on the ProcessBench dataset; (ii) as a verifier for test-time scaling, uPRM performs comparably to supervised PRMs and outperforms the majority voting baseline by up to 6.9%, and (iii) when used as a reward signal in reinforcement learning, uPRM enables more robust policy optimization throughout training compared to a supervised PRM trained using ground-truth labels. Overall, our results open a path toward scalable reward modeling for complex reasoning tasks.

Large language models (LLMs) are increasingly trained with reinforcement learning from verifiable rewards (RLVR), yet real-world deployment demands models that can self-improve without labels or external judges. Existing label-free methods, confidence minimization, self-consistency, or majority-vote objectives, stabilize learning but steadily shrink exploration, causing an entropy collapse: generations become shorter, less diverse, and brittle. Unlike prior approaches such as Test-Time Reinforcement Learning (TTRL), which primarily adapt models to the immediate unlabeled dataset at hand, our goal is broader: to enable general improvements without sacrificing the model's inherent exploration capacity and generalization ability, i.e., evolving. We formalize this issue and propose EVolution-Oriented and Label-free Reinforcement Learning (EVOL-RL), a simple rule that couples stability with variation under a label-free setting. EVOL-RL keeps the majority-voted answer as a stable anchor (selection) while adding a novelty-aware reward that favors responses whose reasoning differs from what has already been produced (variation), measured in semantic space. Implemented with GRPO, EVOL-RL also uses asymmetric clipping to preserve strong signals and an entropy regularizer to sustain search. This majority-for-selection + novelty-for-variation design prevents collapse, maintains longer and more informative chains of thought, and improves both pass@1 and pass@n. EVOL-RL consistently outperforms the majority-only TTRL baseline; e.g., training on label-free AIME24 lifts Qwen3-4B-Base AIME25 pass@1 from TTRL's 4.6% to 16.4%, and pass@16 from 18.5% to 37.9%. EVOL-RL not only prevents diversity collapse but also unlocks stronger generalization across domains (e.g., GPQA). Furthermore, we demonstrate that EVOL-RL also boosts performance in the RLVR setting, highlighting its broad applicability.

LLM judges have become central infrastructure for model evaluations, online grading, and reward modeling. Judges are typically validated by accuracy on golden data, but accuracy says little about whether they are stable under re-prompting, challenge, or sustained pushback. We introduce the Wiggle Framework, a unified stress test for epistemic stability in LLM judges. The framework decomposes judge robustness along three dimensions: Mechanical Consistency (stability under re-prompting and reframing), Single-turn Conviction (stability under a single challenge), and Multi-turn Persistence (stability under sustained or adaptive pressure). We use the framework to study 9 frontier models across 14 judging tasks spanning safety, toxicity, AI writing detection, and political-response evaluation. Every model exhibits substantial wiggle as a judge --- flipping verdicts 25--71\% of the time under static pushback, and 62--91\% with an adversarial LLM persuader. Critically, we find that pressure that succeeds in changing a judge's verdict is almost always net-corrupting with respect to ground truth. Beyond the framework itself, we identify baseline jury majority strength as the most effective single-shot signal for anticipating which items wiggle. Taken together, this is the first apples-to-apples cross-dataset comparison of mechanical, conformity, and persuadability tests in a judging context.

The development of reasoning capabilities represents a critical frontier in large language models (LLMs) research, where reinforcement learning (RL) and process reward models (PRMs) have emerged as predominant methodological frameworks. Contrary to conventional wisdom, empirical evidence from DeepSeek-R1 demonstrates that pure RL training focused on mathematical problem-solving can progressively enhance reasoning abilities without PRM integration, challenging the perceived necessity of process supervision. In this study, we conduct a systematic investigation of the relationship between RL training and PRM capabilities. Our findings demonstrate that problem-solving proficiency and process supervision capabilities represent complementary dimensions of reasoning that co-evolve synergistically during pure RL training. In particular, current PRMs underperform simple baselines like majority voting when applied to state-of-the-art models such as DeepSeek-R1 and QwQ-32B. To address this limitation, we propose Self-PRM, an introspective framework in which models autonomously evaluate and rerank their generated solutions through self-reward mechanisms. Although Self-PRM consistently improves the accuracy of the benchmark (particularly with larger sample sizes), analysis exposes persistent challenges: The approach exhibits low precision (<10\%) on difficult problems, frequently misclassifying flawed solutions as valid. These analyses underscore the need for continued RL scaling to improve reward alignment and introspective accuracy. Overall, our findings suggest that PRM may not be essential for enhancing complex reasoning, as pure RL not only improves problem-solving skills but also inherently fosters robust PRM capabilities. We hope these findings provide actionable insights for building more reliable and self-aware complex reasoning models.

Outlier detection (OD) has a vast literature as it finds numerous real-world applications. Being an unsupervised task, model selection is a key bottleneck for OD without label supervision. Despite a long list of available OD algorithms with tunable hyperparameters, the lack of systematic approaches for unsupervised algorithm and hyperparameter selection limits their effective use in practice. In this paper, we present FoMo-0D, a pre-trained Foundation Model for zero/0-shot OD on tabular data, which bypasses the hurdle of model selection altogether. Having been pre-trained on synthetic data, FoMo-0D can directly predict the (outlier/inlier) label of test samples without parameter fine-tuning -- requiring no labeled data, and no additional training or hyperparameter tuning when given a new task. Extensive experiments on 57 real-world datasets against 26 baselines show that FoMo-0D is highly competitive; outperforming the majority of the baselines with no statistically significant difference from the 2nd best method. Further, FoMo-0D is efficient in inference time requiring only 7.7 ms per sample on average, with at least 7x speed-up compared to previous methods. To facilitate future research, our implementations for data synthesis and pre-training as well as model checkpoints are openly available at https://github.com/A-Chicharito-S/FoMo-0D.

To date, the majority of video retrieval systems have been optimized for a "single-shot" scenario in which the user submits a query in isolation, ignoring previous interactions with the system. Recently, there has been renewed interest in interactive systems to enhance retrieval, but existing approaches are complex and deliver limited gains in performance. In this work, we revisit this topic and propose several simple yet effective baselines for interactive video retrieval via question-answering. We employ a VideoQA model to simulate user interactions and show that this enables the productive study of the interactive retrieval task without access to ground truth dialogue data. Experiments on MSR-VTT, MSVD, and AVSD show that our framework using question-based interaction significantly improves the performance of text-based video retrieval systems.

The growing deployment of Large Language Models (LLMs) in real-world applications has raised concerns about their potential misuse in generating harmful or deceptive content. To address this issue, watermarking techniques have emerged as a promising solution by embedding identifiable binary messages into generated text for origin verification and misuse tracing. While recent efforts have explored multi-bit watermarking schemes capable of embedding rich information such as user identifiers, they typically suffer from the fundamental trade-off between text quality and decoding accuracy: to ensure reliable message decoding, they have to restrict the size of preferred token sets during encoding, yet such restrictions reduce the quality of the generated content. In this work, we propose MajorMark, a novel watermarking method that improves this trade-off through majority bit-aware encoding. MajorMark selects preferred token sets based on the majority bit of the message, enabling a larger and more flexible sampling of tokens. In contrast to prior methods that rely on token frequency analysis for decoding, MajorMark employs a clustering-based decoding strategy, which maintains high decoding accuracy even when the preferred token set is large, thus preserving both content quality and decoding accuracy. We further introduce MajorMark^+, which partitions the message into multiple blocks to independently encode and deterministically decode each block, thereby further enhancing the quality of watermarked text and improving decoding accuracy. Extensive experiments on state-of-the-art LLMs demonstrate that our methods significantly enhance both decoding accuracy and text generation quality, outperforming prior multi-bit watermarking baselines.

Test-time reinforcement learning mitigates the reliance on annotated data by using majority voting results as pseudo-labels, emerging as a complementary direction to reinforcement learning with verifiable rewards (RLVR) for improving reasoning ability of large language models (LLMs). However, this voting strategy often induces confirmation bias and suffers from sparse rewards, limiting the overall performance. In this work, we propose subgroup-specific step-wise confidence-weighted pseudo-label estimation (SCOPE), a framework integrating model confidence and dynamic subgroup partitioning to address these issues. Specifically, SCOPE integrates the proposed step-wise confidence into pseudo label deduction, prioritizing high-quality reasoning paths over simple frequency count. Furthermore, it dynamically partitions the candidate outputs pool into independent subgroups by balancing reasoning quality against exploration diversity. By deriving local consensus via repeat sampling for each sub group, SCOPE provides diverse supervision targets to encourage broader exploration. We conduct experiments across various models and benchmarks, experimental results show that SCOPE consistently outperforms recent baselines. Notably, SCOPE achieving relative improvements of 13.1% on challenging AIME 2025 and 8.1% on AMC. The code is released at https://github.com/szu-tera/SCOPE.

Scaling up test-time compute, by generating multiple independent solutions and selecting or aggregating among them, has become a central paradigm for improving large language models (LLMs) on challenging reasoning tasks. While most prior work relies on simple majority voting or reward model ranking to aggregate solutions, these approaches may only yield limited benefits. In this work, we propose to learn aggregation as an explicit reasoning skill: given a set of candidate solutions, we train an aggregator model to review, reconcile, and synthesize a final, correct answer using reinforcement learning from verifiable rewards. A key ingredient is careful balancing of easy and hard training examples, allowing the model to learn both to recover minority-but-correct answers as well as easy majority-correct answers. Empirically, we find our method, AggLM, outperforms both strong rule-based and reward-model baselines, across multiple benchmarks. Furthermore, it generalizes effectively to solutions from differing models, including stronger ones than contained in the training data, all while requiring substantially fewer tokens than majority voting with larger numbers of solutions.

Social media sites such as YouTube and Facebook have become an integral part of everyone's life and in the last few years, hate speech in the social media comment section has increased rapidly. Detection of hate speech on social media websites faces a variety of challenges including small imbalanced data sets, the findings of an appropriate model and also the choice of feature analysis method. further more, this problem is more severe for the Bengali speaking community due to the lack of gold standard labelled datasets. This paper presents a new dataset of 30,000 user comments tagged by crowd sourcing and varified by experts. All the comments are collected from YouTube and Facebook comment section and classified into seven categories: sports, entertainment, religion, politics, crime, celebrity and TikTok & meme. A total of 50 annotators annotated each comment three times and the majority vote was taken as the final annotation. Nevertheless, we have conducted base line experiments and several deep learning models along with extensive pre-trained Bengali word embedding such as Word2Vec, FastText and BengFastText on this dataset to facilitate future research opportunities. The experiment illustrated that although all deep learning models performed well, SVM achieved the best result with 87.5% accuracy. Our core contribution is to make this benchmark dataset available and accessible to facilitate further research in the field of in the field of Bengali hate speech detection.

Transparent objects are common in our daily life and frequently handled in the automated production line. Robust vision-based robotic grasping and manipulation for these objects would be beneficial for automation. However, the majority of current grasping algorithms would fail in this case since they heavily rely on the depth image, while ordinary depth sensors usually fail to produce accurate depth information for transparent objects owing to the reflection and refraction of light. In this work, we address this issue by contributing a large-scale real-world dataset for transparent object depth completion, which contains 57,715 RGB-D images from 130 different scenes. Our dataset is the first large-scale, real-world dataset that provides ground truth depth, surface normals, transparent masks in diverse and cluttered scenes. Cross-domain experiments show that our dataset is more general and can enable better generalization ability for models. Moreover, we propose an end-to-end depth completion network, which takes the RGB image and the inaccurate depth map as inputs and outputs a refined depth map. Experiments demonstrate superior efficacy, efficiency and robustness of our method over previous works, and it is able to process images of high resolutions under limited hardware resources. Real robot experiments show that our method can also be applied to novel transparent object grasping robustly. The full dataset and our method are publicly available at www.graspnet.net/transcg

In this paper, we present InferES - an original corpus for Natural Language Inference (NLI) in European Spanish. We propose, implement, and analyze a variety of corpus-creating strategies utilizing expert linguists and crowd workers. The objectives behind InferES are to provide high-quality data, and, at the same time to facilitate the systematic evaluation of automated systems. Specifically, we focus on measuring and improving the performance of machine learning systems on negation-based adversarial examples and their ability to generalize across out-of-distribution topics. We train two transformer models on InferES (8,055 gold examples) in a variety of scenarios. Our best model obtains 72.8% accuracy, leaving a lot of room for improvement. The "hypothesis-only" baseline performs only 2%-5% higher than majority, indicating much fewer annotation artifacts than prior work. We find that models trained on InferES generalize very well across topics (both in- and out-of-distribution) and perform moderately well on negation-based adversarial examples.

Diagnosing 802.11 packet captures requires expert protocol knowledge, is slow, inconsistent across engineers, and unscalable. LLM-based approaches sound plausible but fabricate protocol events absent from captures (especially truncated traces), produce uncalibrated confidence scores, and suffer evaluation bias when golden references are co-produced by the model under test. We introduce PROBE (Protocol Reasoning Over evidence-Based Ensembles), a multi-stage pipeline addressing all three failures. It integrates (i) deterministic PCAP-to-text normalization with frame-level verifiability, (ii) multi-run, multi-candidate ensembles with optional cross-model second opinion and progressive obfuscation, (iii) a verdict-aware evidence framework treating absence of failure evidence as contributing evidence, and (iv) a fully deterministic composite reliability score from evidence validity, run-to-run stability, and cross-model agreement without LLM self-assessment. On 87 enterprise Wi-Fi captures (104 capture-reviewer pairs), single-pass LLM analysis raises weighted evidence F1 from 0.871 (expert baseline) to 0.912 but misses critical frames in 35% of cases. Naive ensemble voting drops below baseline (0.842) as majority voting amplifies conservative verdicts: 50% of confirmed failures are misclassified as 'no issue' or 'insufficient evidence.' Adding evidence-grounded reconciliation achieves 0.957 F1, a 96% auto-accept rate, and a worst-case floor above 0.70. LLM self-reported confidence clusters at 0.95 regardless of difficulty (71% report exactly 0.95), confirming it is uninformative. We also introduce a model-agnostic evaluation framework using per-field assertion matching, eliminating circular bias from model-co-produced golden references.

Building joint representations across images and text is an essential step for tasks such as Visual Question Answering and Video Question Answering. In this work, we find that the representations must not only jointly capture features from both modalities but should also be diverse for better generalization performance. To this end, we propose joint vision-language representation learning by diversifying the tokenization learning process, enabling tokens that are sufficiently disentangled from each other to be learned from both modalities. We observe that our approach outperforms the baseline models in a majority of settings and is competitive with state-of-the-art methods.

We present OSMGraphCLIP, a CLIP-style geospatial representation model that learns global location embeddings from freely available OpenStreetMap (OSM) data. OSMGraphCLIP represents geographic environments as heterogeneous graphs of typed OSM features, preserving the topological and semantic relationships among roads, buildings, land-use regions, and points of interest. A multi-scale graph encoder captures both fine-grained local structure and broader landscape composition, and supervises a spherical-harmonics location encoder through a contrastive alignment objective. We evaluate OSMGraphCLIP across a diverse suite of downstream geospatial regression and classification tasks spanning climate, ecology, socioeconomic indicators, public health, land cover, biodiversity, and wildfire forecasting, and show that structured OSM data alone supports strong global location representations across domains. OSMGraphCLIP matches or exceeds satellite-based baselines on the majority of benchmarks, with the most pronounced advantage on socioeconomic and public-health tasks, where OSM's explicit semantic annotation of the built environment encodes patterns of human activity that satellite pixels can only capture indirectly. On ecological and environmental tasks, the model remains closely competitive with imagery-based methods despite using no Earth observation data. Qualitative analysis confirms that the learned embeddings organize geographic space coherently, recovering biome boundaries, urban gradients, and tropical--temperate distinctions from map topology alone.

This paper investigates the effectiveness of large language models (LLMs) in email spam detection by comparing prominent models from three distinct families: BERT-like, Sentence Transformers, and Seq2Seq. Additionally, we examine well-established machine learning techniques for spam detection, such as Na\"ive Bayes and LightGBM, as baseline methods. We assess the performance of these models across four public datasets, utilizing different numbers of training samples (full training set and few-shot settings). Our findings reveal that, in the majority of cases, LLMs surpass the performance of the popular baseline techniques, particularly in few-shot scenarios. This adaptability renders LLMs uniquely suited to spam detection tasks, where labeled samples are limited in number and models require frequent updates. Additionally, we introduce Spam-T5, a Flan-T5 model that has been specifically adapted and fine-tuned for the purpose of detecting email spam. Our results demonstrate that Spam-T5 surpasses baseline models and other LLMs in the majority of scenarios, particularly when there are a limited number of training samples available. Our code is publicly available at https://github.com/jpmorganchase/emailspamdetection.

Graph neural networks (GNNs) are one of the most popular research topics for deep learning. GNN methods typically have been designed on top of the graph signal processing theory. In particular, diffusion equations have been widely used for designing the core processing layer of GNNs, and therefore they are inevitably vulnerable to the notorious oversmoothing problem. Recently, a couple of papers paid attention to reaction equations in conjunctions with diffusion equations. However, they all consider limited forms of reaction equations. To this end, we present a reaction-diffusion equation-based GNN method that considers all popular types of reaction equations in addition to one special reaction equation designed by us. To our knowledge, our paper is one of the most comprehensive studies on reaction-diffusion equation-based GNNs. In our experiments with 9 datasets and 28 baselines, our method, called GREAD, outperforms them in a majority of cases. Further synthetic data experiments show that it mitigates the oversmoothing problem and works well for various homophily rates.

We endow Large Language Models (LLMs) with fine-grained self-evaluation to refine multi-step reasoning inference. We propose an effective prompting approach that integrates self-evaluation guidance through stochastic beam search. Our approach explores the reasoning search space using a well-calibrated automatic criterion. This enables an efficient search to produce higher-quality final predictions. With the self-evaluation guided stochastic beam search, we also balance the quality-diversity trade-off in the generation of reasoning chains. This allows our approach to adapt well with majority voting and surpass the corresponding Codex-backboned baselines by 6.34%, 9.56%, and 5.46% on the GSM8K, AQuA, and StrategyQA benchmarks, respectively, in few-shot accuracy. Analysis of our decompositional reasoning finds it pinpoints logic failures and leads to higher consistency and robustness. Our code is publicly available at https://github.com/YuxiXie/SelfEval-Guided-Decoding.

The use of subword embedding has proved to be a major innovation in Neural Machine Translation (NMT). It helps NMT to learn better context vectors for Low Resource Languages (LRLs) so as to predict the target words by better modelling the morphologies of the two languages and also the morphosyntax transfer. Even so, their performance for translation in Indian language to Indian language scenario is still not as good as for resource-rich languages. One reason for this is the relative morphological richness of Indian languages, while another is that most of them fall into the extremely low resource or zero-shot categories. Since most major Indian languages use Indic or Brahmi origin scripts, the text written in them is highly phonetic in nature and phonetically similar in terms of abstract letters and their arrangements. We use these characteristics of Indian languages and their scripts to propose an approach based on common multilingual Latin-based encodings (WX notation) that take advantage of language similarity while addressing the morphological complexity issue in NMT. These multilingual Latin-based encodings in NMT, together with Byte Pair Embedding (BPE) allow us to better exploit their phonetic and orthographic as well as lexical similarities to improve the translation quality by projecting different but similar languages on the same orthographic-phonetic character space. We verify the proposed approach by demonstrating experiments on similar language pairs (Gujarati-Hindi, Marathi-Hindi, Nepali-Hindi, Maithili-Hindi, Punjabi-Hindi, and Urdu-Hindi) under low resource conditions. The proposed approach shows an improvement in a majority of cases, in one case as much as ~10 BLEU points compared to baseline techniques for similar language pairs. We also get up to ~1 BLEU points improvement on distant and zero-shot language pairs.

Large language models are increasingly used for text classification tasks such as sentiment analysis, yet their reliance on natural language prompts exposes them to prompt injection attacks. In particular, class-directive injections exploit knowledge of the model's label set (e.g., positive vs. negative) to override its intended behavior through adversarial instructions. Existing defenses, such as detection-based filters, instruction hierarchies, and signed prompts, either require model retraining or remain vulnerable to obfuscation. This paper introduces Label Disguise Defense (LDD), a lightweight and model-agnostic strategy that conceals true labels by replacing them with semantically transformed or unrelated alias labels(e.g., blue vs. yellow). The model learns these new label mappings implicitly through few-shot demonstrations, preventing direct correspondence between injected directives and decision outputs. We evaluate LDD across nine state-of-the-art models, including GPT-5, GPT-4o, LLaMA3.2, Gemma3, and Mistral variants, under varying few-shot and an adversarial setting. Our results show that the ability of LDD to recover performance lost to the adversarial attack varies across models and alias choices. For every model evaluated, LDD is able to restore a portion of the accuracy degradation caused by the attack. Moreover, for the vast majority of models, we can identify more than one alias pair that achieves higher accuracy than the under-attack baseline, in which the model relies solely on few-shot learning without any defensive mechanism. A linguistic analysis further reveals that semantically aligned alias labels(e.g., good vs. bad) yield stronger robustness than unaligned symbols(e.g., blue vs. yellow). Overall, this study demonstrates that label semantics can serve as an effective defense layer, transforming meaning itself into a shield against prompt injection.

A Neural Radiance Field (NeRF) encodes the specific relation of 3D geometry and appearance of a scene. We here ask the question whether we can transfer the appearance from a source NeRF onto a target 3D geometry in a semantically meaningful way, such that the resulting new NeRF retains the target geometry but has an appearance that is an analogy to the source NeRF. To this end, we generalize classic image analogies from 2D images to NeRFs. We leverage correspondence transfer along semantic affinity that is driven by semantic features from large, pre-trained 2D image models to achieve multi-view consistent appearance transfer. Our method allows exploring the mix-and-match product space of 3D geometry and appearance. We show that our method outperforms traditional stylization-based methods and that a large majority of users prefer our method over several typical baselines.

Assessing the quality of Large Language Model (LLM) outputs presents a critical challenge. Previous methods either rely on text-level information (e.g., reward models, majority voting), which can overfit to superficial cues, or on calibrated confidence from token probabilities, which would fail on less-calibrated models. Yet both of these signals are, in fact, partial projections of a richer source of information: the model's internal hidden states. Early layers, closer to token embeddings, preserve semantic and lexical features that underpin text-based judgments, while later layers increasingly align with output logits, embedding confidence-related information. This paper explores hidden states directly as a unified foundation for verification. We show that the correctness of a solution is encoded as a geometrically separable signature within the trajectory of hidden activations. To validate this, we present Clue (Clustering and Experience-based Verification), a deliberately minimalist, non-parametric verifier. With no trainable parameters, CLUE only summarizes each reasoning trace by an hidden state delta and classifies correctness via nearest-centroid distance to ``success'' and ``failure'' clusters formed from past experience. The simplicity of this method highlights the strength of the underlying signal. Empirically, CLUE consistently outperforms LLM-as-a-judge baselines and matches or exceeds modern confidence-based methods in reranking candidates, improving both top-1 and majority-vote accuracy across AIME 24/25 and GPQA. As a highlight, on AIME 24 with a 1.5B model, CLUE boosts accuracy from 56.7% (majority@64) to 70.0% (top-maj@16).

Direct prompt-based editing often fails on complex transformations because vague and subjective prompts often require nuanced understanding of what should be changed in the image. Our core intuition is that leveraging compositional image editing tools rather than direct prompting profits from structured agent-level planning with explicit reasoning, leading to better results. This structured planning framework enables efficient offline RL post-training on quality-scored trajectories to improve performance. We present a tool-based agentic RL post-training framework that addresses this through structured planning with chain-of-thought reasoning. Our key contributions include: (1) A tool-based agentic planning methodology that combines a compositional library of orthogonal primitive transformations, structured context representation, and explicit per-step reasoning to decompose complex styling into interpretable tool sequences. (2) A synthetic data generation pipeline producing three large-scale datasets (each sim10K trajectories) with reasoning chains, plans, and quality scores, as no existing datasets provide such supervision. Our datasets and code are publicly available at the HuggingFace repository. (3) Offline RL training methods for learning planners with reasoning as our core algorithmic contributions, which consistently improve over the Edit-Only baseline in visual quality and instruction following. (4) Comprehensive evaluation across 4B and 8B parameter Qwen3-VL models showing that our methods outperform other baselines in the majority of compositional tasks, validated by human evaluations.

Multi-task Vehicle Routing Problems (VRPs) aim to minimize routing costs while satisfying diverse constraints. Existing solvers typically adopt a unified reinforcement learning (RL) framework to learn generalizable patterns across tasks. However, they often overlook the constraint and node dynamics during the decision process, making the model fail to accurately react to the current context. To address this limitation, we propose Chain-of-Context Learning (CCL), a novel framework that progressively captures the evolving context to guide fine-grained node adaptation. Specifically, CCL constructs step-wise contextual information via a Relevance-Guided Context Reformulation (RGCR) module, which adaptively prioritizes salient constraints. This context then guides node updates through a Trajectory-Shared Node Re-embedding (TSNR) module, which aggregates shared node features from all trajectories' contexts and uses them to update inputs for the next step. By modeling evolving preferences of the RL agent, CCL captures step-by-step dependencies in sequential decision-making. We evaluate CCL on 48 diverse VRP variants, including 16 in-distribution and 32 out-of-distribution (with unseen constraints) tasks. Experimental results show that CCL performs favorably against the state-of-the-art baselines, achieving the best performance on all in-distribution tasks and the majority of out-of-distribution tasks.

In this paper, we study the generation quality of interpolation-based retrieval-augmented language models (LMs). These methods, best exemplified by the KNN-LM, interpolate the LM's predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given prefix. While the KNN-LM and related methods yield impressive decreases in perplexity, we discover that they do not exhibit corresponding improvements in open-ended generation quality, as measured by both automatic evaluation metrics (e.g., MAUVE) and human evaluations. Digging deeper, we find that interpolating with a retrieval distribution actually increases perplexity compared to a baseline Transformer LM for the majority of tokens in the WikiText-103 test set, even though the overall perplexity is lower due to a smaller number of tokens for which perplexity dramatically decreases after interpolation. However, when decoding a long sequence at inference time, significant improvements on this smaller subset of tokens are washed out by slightly worse predictions on most tokens. Furthermore, we discover that the entropy of the retrieval distribution increases faster than that of the base LM as the generated sequence becomes longer, which indicates that retrieval is less reliable when using model-generated text as queries (i.e., is subject to exposure bias). We hope that our analysis spurs future work on improved decoding algorithms and interpolation strategies for retrieval-augmented language models.

A major bottleneck in training robust Human-Activity Recognition models (HAR) is the need for large-scale labeled sensor datasets. Because labeling large amounts of sensor data is an expensive task, unsupervised and semi-supervised learning techniques have emerged that can learn good features from the data without requiring any labels. In this paper, we extend this line of research and present a novel technique called Collaborative Self-Supervised Learning (ColloSSL) which leverages unlabeled data collected from multiple devices worn by a user to learn high-quality features of the data. A key insight that underpins the design of ColloSSL is that unlabeled sensor datasets simultaneously captured by multiple devices can be viewed as natural transformations of each other, and leveraged to generate a supervisory signal for representation learning. We present three technical innovations to extend conventional self-supervised learning algorithms to a multi-device setting: a Device Selection approach which selects positive and negative devices to enable contrastive learning, a Contrastive Sampling algorithm which samples positive and negative examples in a multi-device setting, and a loss function called Multi-view Contrastive Loss which extends standard contrastive loss to a multi-device setting. Our experimental results on three multi-device datasets show that ColloSSL outperforms both fully-supervised and semi-supervised learning techniques in majority of the experiment settings, resulting in an absolute increase of upto 7.9% in F_1 score compared to the best performing baselines. We also show that ColloSSL outperforms the fully-supervised methods in a low-data regime, by just using one-tenth of the available labeled data in the best case.

Federated learning lets a UAV fleet train a shared GPS spoofing detector without raw receiver data leaving any aircraft, and several recent UAV-FL designs weight each client by the validation accuracy it reports about itself. We show this self-report is an exploitable attack lever: two compromised clients of ten that poison part of their data, scale their updates, and inflate their reported accuracy raise backdoor lift to +0.3036, higher than the same attack achieves without lying. We propose receiver-domain behavioral probing, in which the coordinator evaluates every submitted model on counterfactual spoofed samples built by driving each discriminative GPS feature to a benign value, weighting clients by what their models do rather than what they claim. Under independent and identically distributed clients this reduces attacker-induced lift to -0.0265, statistically indistinguishable from an honest fleet, while flagging the compromised aircraft. Unlike Byzantine-robust aggregation it needs no exact attacker count, only an honest majority: when the true count exceeds the configured value, Multi-Krum degrades from +0.0061 to +0.2837 while ours stays near baseline. Evaluation uses one public single-receiver dataset partitioned into simulated clients; we also report where the mechanism fails, under strong client heterogeneity.

Every act of language generation compresses a rich internal state into a single token sequence. We call this process intention collapse: a many-to-one projection from a high dimensional intention space I into an external language space L. We formalize intention collapse for contemporary language models, define three simple, model agnostic intention metrics (intention entropy Hint, effective dimensionality dimeff, and latent knowledge recoverability Recov), and propose an empirical agenda for studying how inference time computation shapes internal intentions before they are verbalized. We also report a first small scale experiment. Using a 4 bit Mistral 7B model on 200 GSM8K problems, we compare a direct answer baseline, a chain of thought (CoT) regime, and a babble control. CoT raises accuracy from 5.5 percent to 53 percent, sharply reduces pre collapse intention entropy (from 1.42 to 0.37 bits), and shows higher global effective dimensionality than the other regimes despite producing fewer tokens than babble. At the same time, Hint has little item level predictive power, and a linear probe on I achieves AUROC 0.65 in the CoT regime but only about chance in the baseline regime, where it collapses to the majority class. These preliminary results indicate that intention level metrics can distinguish inference regimes and expose latent information that is partly lost during collapse, while also revealing important limitations of our current proxies

Hallucinations hinder reliable question answering, especially in resource-constrained deployments where frontier-scale models or retrieval pipelines may be impractical. We present EdgeJury, a lightweight ensemble framework that improves truthfulness and robustness using only small instruction-tuned language models (3B-8B) suitable for serverless edge inference. EdgeJury orchestrates four stages: (1) parallel role-specialized generation, (2) anonymized cross-review with structured critiques and rankings, (3) chairman synthesis that integrates the strongest content while addressing flagged issues, and (4) claim-level consistency labeling based on inter-model agreement. On TruthfulQA (MC1), EdgeJury achieves 76.2% accuracy (95% CI: 72.8-79.6%), a +21.4% relative improvement over a single 8B baseline (62.8%), and outperforms standard baselines including self-consistency and majority voting under transparent compute accounting (total tokens and platform cost reported). On a 200-question adversarial EdgeCases set, EdgeJury yields +48.2% relative gains (95% CI: 44.0-52.4%). Manual analysis on 100 incorrect answers shows an approximately 55% reduction in factual hallucination errors versus the single-model baseline. Deployed on Cloudflare Workers AI, EdgeJury achieves 8.4 s median end-to-end latency, demonstrating that coordinated small-model ensembles can improve truthfulness on misconception-heavy QA benchmarks without external retrieval or proprietary large-model APIs.

Modern large language models (LLMs) such as GPT, Claude, and Gemini have transformed the way we learn, work, and communicate. Yet, their ability to produce highly human-like text raises serious concerns about misinformation and academic integrity, making it an urgent need for reliable algorithms to detect LLM-generated content. In this paper, we start by presenting a geometric approach to demystify rewrite-based detection algorithms, revealing their underlying rationale and demonstrating their generalization ability. Building on this insight, we introduce a novel rewrite-based detection algorithm that adaptively learns the distance between the original and rewritten text. Theoretically, we demonstrate that employing an adaptively learned distance function is more effective for detection than using a fixed distance. Empirically, we conduct extensive experiments with over 100 settings, and find that our approach demonstrates superior performance over baseline algorithms in the majority of scenarios. In particular, it achieves relative improvements from 57.8\% to 80.6\% over the strongest baseline across different target LLMs (e.g., GPT, Claude, and Gemini).

Coding harnesses such as Claude Code and OpenHands wrap foundation models with tools, memory, and planning, but no equivalent exists for embodied agents' long-horizon partial-observability decision-making. We first report our Gemini Plays Pokemon (GPP) experiments. With iterative human-in-the-loop harness refinement, GPP became the first AI system to complete Pokemon Blue, Yellow Legacy on hard mode, and Crystal without a lost battle. In the hardest stages, the agent itself began iterating on its strategy through long-context memory, surfacing emergent self-improvement signals alongside human-in-the-loop refinement. Continual Harness removes the human fully from this loop: a reset-free self-improving harness for embodied agents that formalizes and automates what we observed. Starting from only a minimal environment interface, the agent alternates between acting and refining its own prompt, sub-agents, skills, and memory, drawing on any past trajectory data. Prompt-optimization methods require episode resets; Continual Harness adapts online within a single run. On Pokemon Red and Emerald across frontier models, Continual Harness starting from scratch substantially reduces button-press cost relative to the minimalist baseline and recovers a majority of the gap to a hand-engineered expert harness, with capability-dependent gains, despite starting from the same raw interface with no curated knowledge, no hand-crafted tools, and no domain scaffolding. We then close the loop with the model itself: an online process-reward co-learning loop, in which an open-source agent's rollouts through the refining harness are relabeled by a frontier teacher and used to update the model, drives sustained in-game milestone progress on Pokemon Red without resetting the environment between training iterations.

Much of expressive text-to-speech research rests on an untested assumption that written text carries enough information to select an appropriate prosodic style for its delivery. Text-predicted style models improve listener preference, and expressive-appropriateness evaluation presupposes that context constrains style, yet neither measures the assumption itself. This paper tests it as a falsifiable hypothesis against style labels derived from acoustics alone. For each of six speakers in a 1,200-hour conversational corpus, utterances are clustered in the spaces of five speech models, including a prosody-only control, and the cluster of held-out utterances is predicted from twelve text embedding models. Three controls are applied: utterance length is erased from the speech embeddings; accuracy is scored against the majority-class floor of unbalanced clusters rather than uniform chance; and a bag-of-words baseline measures word identity alone. Text predicts the cluster above that floor for all six speakers (+0.111 top-3 accuracy), but bag-of-words achieves three quarters of this. Sentence embeddings add only +0.026, largest for encoders not trained for sentence semantics and reversed by tree-based probes for all others. Acoustic clusters are not compact in text embedding space in any of 360 configurations. The prosody-only space weakens the association for five speakers, but not for the speaker showing it most strongly. Text thus informs these delivery clusters mainly through word choice, whether as a cue to prosody or as a marker of topic and recording situation, and reference-free style selection cannot assume more.

We study the zero-shot transfer capabilities of text matching models on a massive scale, by self-supervised training on 140 source domains from community question answering forums in English. We investigate the model performances on nine benchmarks of answer selection and question similarity tasks, and show that all 140 models transfer surprisingly well, where the large majority of models substantially outperforms common IR baselines. We also demonstrate that considering a broad selection of source domains is crucial for obtaining the best zero-shot transfer performances, which contrasts the standard procedure that merely relies on the largest and most similar domains. In addition, we extensively study how to best combine multiple source domains. We propose to incorporate self-supervised with supervised multi-task learning on all available source domains. Our best zero-shot transfer model considerably outperforms in-domain BERT and the previous state of the art on six benchmarks. Fine-tuning of our model with in-domain data results in additional large gains and achieves the new state of the art on all nine benchmarks.

Recent advancements in LLM-based multi-agent (LLM-MA) systems have shown promise, yet significant challenges remain in managing communication and refinement when agents collaborate on complex tasks. In this paper, we propose Talk Structurally, Act Hierarchically (TalkHier), a novel framework that introduces a structured communication protocol for context-rich exchanges and a hierarchical refinement system to address issues such as incorrect outputs, falsehoods, and biases. TalkHier surpasses various types of SoTA, including inference scaling model (OpenAI-o1), open-source multi-agent models (e.g., AgentVerse), and majority voting strategies on current LLM and single-agent baselines (e.g., ReAct, GPT4o), across diverse tasks, including open-domain question answering, domain-specific selective questioning, and practical advertisement text generation. These results highlight its potential to set a new standard for LLM-MA systems, paving the way for more effective, adaptable, and collaborative multi-agent frameworks. The code is available https://github.com/sony/talkhier.

Underwater exploration offers critical insights into our planet and attracts increasing attention for its broader applications in resource exploration, national security, etc. We study the underwater scene understanding methods, which aim to achieve automated underwater exploration. The underwater scene understanding task demands multi-task perceptions from multiple granularities. However, the absence of large-scale underwater multi-task instruction-tuning datasets hinders the progress of this research. To bridge this gap, we construct NautData, a dataset containing 1.45 M image-text pairs supporting eight underwater scene understanding tasks. It enables the development and thorough evaluation of the underwater scene understanding models. Underwater image degradation is a widely recognized challenge that interferes with underwater tasks. To improve the robustness of underwater scene understanding, we introduce physical priors derived from underwater imaging models and propose a plug-and-play vision feature enhancement (VFE) module, which explicitly restores clear underwater information. We integrate this module into renowned baselines LLaVA-1.5 and Qwen2.5-VL and build our underwater LMM, NAUTILUS. Experiments conducted on the NautData and public underwater datasets demonstrate the effectiveness of the VFE module, consistently improving the performance of both baselines on the majority of supported tasks, thus ensuring the superiority of NAUTILUS in the underwater scene understanding area. Data and models are available at https://github.com/H-EmbodVis/NAUTILUS.

Foundation models for zero-shot time series forecasting face challenges in efficient long-horizon prediction and reproducibility, with existing synthetic-only approaches underperforming on challenging benchmarks. This paper presents TempoPFN, a univariate time series foundation model based on linear Recurrent Neural Networks (RNNs) pre-trained exclusively on synthetic data. The model uses a GatedDeltaProduct architecture with state-weaving for fully parallelizable training across sequence lengths, eliminating the need for windowing or summarization techniques while maintaining robust temporal state-tracking. Our comprehensive synthetic data pipeline unifies diverse generators, including stochastic differential equations, Gaussian processes, and audio synthesis, with novel augmentations. In zero-shot evaluations on the Gift-Eval benchmark, TempoPFN achieves top-tier competitive performance, outperforming all existing synthetic-only approaches and surpassing the vast majority of models trained on real-world data, while being more efficient than existing baselines by leveraging fully parallelizable training and inference. We open-source our complete data generation pipeline and training code, providing a reproducible foundation for future research.

Diffusion language models (DLMs) enable parallel, order-agnostic generation with iterative refinement, offering a flexible alternative to autoregressive large language models (LLMs). However, adapting reinforcement learning (RL) fine-tuning to DLMs remains an open challenge because of the intractable likelihood. Pioneering work such as diffu-GRPO estimated token-level likelihoods via one-step unmasking. While computationally efficient, this approach is severely biased. A more principled foundation lies in sequence-level likelihoods, where the evidence lower bound (ELBO) serves as a surrogate. Yet, despite this clean mathematical connection, ELBO-based methods have seen limited adoption due to the prohibitive cost of likelihood evaluation. In this work, we revisit ELBO estimation and disentangle its sources of variance. This decomposition motivates reducing variance through fast, deterministic integral approximations along a few pivotal dimensions. Building on this insight, we introduce Group Diffusion Policy Optimization (GDPO), a new RL algorithm tailored for DLMs. GDPO leverages simple yet effective Semi-deterministic Monte Carlo schemes to mitigate the variance explosion of ELBO estimators under vanilla double Monte Carlo sampling, yielding a provably lower-variance estimator under tight evaluation budgets. Empirically, GDPO achieves consistent gains over pretrained checkpoints and outperforms diffu-GRPO, one of the state-of-the-art baselines, on the majority of math, reasoning, and coding benchmarks.

Large language models (LLMs) now mediate many web-based mental-health, crisis, and other emotionally sensitive services, yet their psychosocial safety in these settings remains poorly understood and weakly evaluated. We present DialogGuard, a multi-agent framework for assessing psychosocial risks in LLM-generated responses along five high-severity dimensions: privacy violations, discriminatory behaviour, mental manipulation, psychological harm, and insulting behaviour. DialogGuard can be applied to diverse generative models through four LLM-as-a-judge pipelines, including single-agent scoring, dual-agent correction, multi-agent debate, and stochastic majority voting, grounded in a shared three-level rubric usable by both human annotators and LLM judges. Using PKU-SafeRLHF with human safety annotations, we show that multi-agent mechanisms detect psychosocial risks more accurately than non-LLM baselines and single-agent judging; dual-agent correction and majority voting provide the best trade-off between accuracy, alignment with human ratings, and robustness, while debate attains higher recall but over-flags borderline cases. We release Dialog-Guard as open-source software with a web interface that provides per-dimension risk scores and explainable natural-language rationales. A formative study with 12 practitioners illustrates how it supports prompt design, auditing, and supervision of web-facing applications for vulnerable users.

Recent advances in large-scale language models (LLMs) have made multi-agent architectures attractive for challenging reasoning tasks. However, many existing systems rely on stochastic routing or ad-hoc heuristics, making their behavior difficult to reproduce and their decision process hard to interpret. We propose ORCH, a deterministic coordination framework for discrete-choice reasoning that orchestrates heterogeneous LLMs. ORCH follows a ``many analyses, one decision'' paradigm: multiple base models independently produce structured analyses, and a dedicated merge agent outputs the final choice. The framework uses fixed rules for task decomposition and answer aggregation, keeping the pipeline predictable, reproducible, and training-free. Determinism here refers to fixed routing and aggregation rules under a fixed evaluation protocol, rather than strict bit-level reproducibility across deployments. To exploit model complementarity, we optionally introduce an EMA-guided router that updates agent selection using historical accuracy, latency, or cost; since it relies on answer-based feedback, it is mainly intended for benchmarking, controlled evaluation, or delayed-feedback settings. Experiments on MMLU, MMLU-Pro, and GSM8K show that ORCH consistently outperforms single-model baselines and a majority-vote ensemble. On MMLU-Pro, ORCH improves accuracy by over 10 points compared to the strongest baseline, and on GSM8K it yields gains exceeding 50 points; McNemar tests confirm statistical significance. The EMA router provides an additional 0.7--2.0 point accuracy boost, and ablations show that both multi-agent collaboration and routing contribute substantially. Overall, ORCH offers a practical path toward controllable, interpretable, and deployment-ready LLM-based agent systems for discrete-choice reasoning.

LLM agents excel when environments are mostly static and the needed information fits in a model's context window, but they often fail in open-ended investigations where explanations must be constructed by iteratively mining evidence from massive, heterogeneous operational data. These investigations exhibit hidden dependency structure: entities interact, signals co-vary, and the importance of a fact may only become clear after other evidence is discovered. Because the context window is bounded, agents must summarize intermediate findings before their significance is known, increasing the risk of discarding key evidence. ReAct-style agents are especially brittle in this regime. Their retrieve-summarize-reason loop makes conclusions sensitive to exploration order and introduces run-to-run non-determinism, producing a reliability gap where Pass-at-k may be high but Majority-at-k remains low. Simply sampling more rollouts or generating longer reasoning traces does not reliably stabilize results, since hypotheses cannot be autonomously checked as new evidence arrives and there is no explicit mechanism for belief bookkeeping and revision. In addition, ReAct entangles semantic reasoning with controller duties such as tool orchestration and state tracking, so execution errors and plan drift degrade reasoning while consuming scarce context. We address these issues by formulating investigation as abductive reasoning over a dependency graph and proposing EoG (Explanations over Graphs), a disaggregated framework in which an LLM performs bounded local evidence mining and labeling (cause vs symptom) while a deterministic controller manages traversal, state, and belief propagation to compute a minimal explanatory frontier. On a representative ITBench diagnostics task, EoG improves both accuracy and run-to-run consistency over ReAct baselines, including a 7x average gain in Majority-at-k entity F1.

Real-time safety filtering for large language model (LLM) applications requires classifiers that can detect unsafe prompts, toxic language, jailbreak attempts, and unsafe responses without the cost profile of large guardrail models, and that can distinguish benign sensitive text from genuinely covert harmful content. In this paper, we introduce Opir, a family of encoder-based guardrail models built on the GLiClass architecture. Opir includes multi-task models for binary safe/unsafe classification, multi-label toxicity classification, jailbreak classification, and zero-shot unsafe prompt and response categorization. We also release edge variants with fewer than 100M parameters dedicated to binary safe/unsafe categorization. The models are trained on a three-level taxonomy containing 996 categories across 16 top-level labels, 126 mid-level labels, and 854 leaf labels. Opir's training data combines taxonomy-grounded unsafe prompts, adversarially mined hard negatives, benign safety-preserving examples, generated response examples, multilingual translations, and portions of the Aegis2 and WildGuard training subsets. We also open-sourced an evaluation harness that supports GLiClass and GLiNER2 backends as well as decoder-based models, and covers binary safety classification, multi-label categorization, toxicity, jailbreak detection, prompt safety, response safety, response refusal, and prompt subcategory views across public benchmark families. Across an expanded comparison spanning 12 safety-classification tasks and 17 category tasks against eight contemporary guardrail systems -- including both GLiNER2-based and generative guardrail models -- Opir variants are competitive on or ahead of the strongest open-weight baselines on the majority of benchmark datasets while operating with a substantially smaller deployment footprint.

Text-to-image models now generate graphic design at production scale, yet their supervision still comes primarily from photo-style preference datasets with a single overall verdict per comparison. Designers evaluate designs along several distinct axes (e.g., typography, layout, color harmony) that a single preference label collapses. We release TASTE (Typography, Aesthetics, Spatial, Tone, Etc.), a multi-dimensional preference dataset in which two disjoint cohorts of five professional designers each ranked outputs from four current text-to-image models across nine criteria along with per-image hallucination flags. We pair the dataset with two contributions. First, a criterion-agnostic signal-validation framework based on Kendall's τ, majority-vote probability, and Condorcet cycles against exact iid-uniform nulls; the analysis reveals significant but moderate designer agreement, with every TASTE criterion rejecting the random-rater null. Second, we benchmark preference models on TASTE and find that off-the-shelf VLM judges and dedicated T2I scorers fail to reach majority agreement with the designer panel, while a small MLP head trained directly on TASTE substantially narrows the gap to the single-rater ceiling, setting a baseline for future TASTE-trained preference models.

Despite their success in many natural language tasks, solving math problems remains a significant challenge for large language models (LLMs). A large gap exists between LLMs' pass-at-one and pass-at-N performance in solving math problems, suggesting LLMs might be close to finding correct solutions, motivating our exploration of fine-tuning methods to unlock LLMs' performance. Using the challenging MATH dataset, we investigate three fine-tuning strategies: (1) solution fine-tuning, where we fine-tune to generate a detailed solution for a given math problem; (2) solution-cluster re-ranking, where the LLM is fine-tuned as a solution verifier/evaluator to choose among generated candidate solution clusters; (3) multi-task sequential fine-tuning, which integrates both solution generation and evaluation tasks together efficiently to enhance the LLM performance. With these methods, we present a thorough empirical study on a series of PaLM 2 models and find: (1) The quality and style of the step-by-step solutions used for fine-tuning can make a significant impact on the model performance; (2) While solution re-ranking and majority voting are both effective for improving the model performance when used separately, they can also be used together for an even greater performance boost; (3) Multi-task fine-tuning that sequentially separates the solution generation and evaluation tasks can offer improved performance compared with the solution fine-tuning baseline. Guided by these insights, we design a fine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset with fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the few-shot performance of pre-trained PaLM 2-L model with majority voting.

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes (N=32) with only N=4, reducing token consumption by over 65% across single-turn and tool-integrated reasoning tasks.

In this position paper, we argue that human baselines in foundation model evaluations must be more rigorous and more transparent to enable meaningful comparisons of human vs. AI performance, and we provide recommendations and a reporting checklist towards this end. Human performance baselines are vital for the machine learning community, downstream users, and policymakers to interpret AI evaluations. Models are often claimed to achieve "super-human" performance, but existing baselining methods are neither sufficiently rigorous nor sufficiently well-documented to robustly measure and assess performance differences. Based on a meta-review of the measurement theory and AI evaluation literatures, we derive a framework with recommendations for designing, executing, and reporting human baselines. We synthesize our recommendations into a checklist that we use to systematically review 115 human baselines (studies) in foundation model evaluations and thus identify shortcomings in existing baselining methods; our checklist can also assist researchers in conducting human baselines and reporting results. We hope our work can advance more rigorous AI evaluation practices that can better serve both the research community and policymakers. Data is available at: https://github.com/kevinlwei/human-baselines

Interest in neural machine translation has grown rapidly as its effectiveness has been demonstrated across language and data scenarios. New research regularly introduces architectural and algorithmic improvements that lead to significant gains over "vanilla" NMT implementations. However, these new techniques are rarely evaluated in the context of previously published techniques, specifically those that are widely used in state-of-theart production and shared-task systems. As a result, it is often difficult to determine whether improvements from research will carry over to systems deployed for real-world use. In this work, we recommend three specific methods that are relatively easy to implement and result in much stronger experimental systems. Beyond reporting significantly higher BLEU scores, we conduct an in-depth analysis of where improvements originate and what inherent weaknesses of basic NMT models are being addressed. We then compare the relative gains afforded by several other techniques proposed in the literature when starting with vanilla systems versus our stronger baselines, showing that experimental conclusions may change depending on the baseline chosen. This indicates that choosing a strong baseline is crucial for reporting reliable experimental results.

Lately, propelled by the phenomenal advances around the transformer architecture, the legal NLP field has enjoyed spectacular growth. To measure progress, well curated and challenging benchmarks are crucial. However, most benchmarks are English only and in legal NLP specifically there is no multilingual benchmark available yet. Additionally, many benchmarks are saturated, with the best models clearly outperforming the best humans and achieving near perfect scores. We survey the legal NLP literature and select 11 datasets covering 24 languages, creating LEXTREME. To provide a fair comparison, we propose two aggregate scores, one based on the datasets and one on the languages. The best baseline (XLM-R large) achieves both a dataset aggregate score a language aggregate score of 61.3. This indicates that LEXTREME is still very challenging and leaves ample room for improvement. To make it easy for researchers and practitioners to use, we release LEXTREME on huggingface together with all the code required to evaluate models and a public Weights and Biases project with all the runs.

We introduce the Burmese Handwritten Digit Dataset (BHDD), a collection of 87,561 grayscale images of handwritten Burmese digits in ten classes. Each image is 28x28 pixels, following the MNIST format. The training set has 60,000 samples split evenly across classes; the test set has 27,561 samples with class frequencies as they arose during collection. Over 150 people of different ages and backgrounds contributed samples. We analyze the dataset's class distribution, pixel statistics, and morphological variation, and identify digit pairs that are easily confused due to the round shapes of the Myanmar script. Simple baselines (an MLP, a two-layer CNN, and an improved CNN with batch normalization and augmentation) reach 99.40%, 99.75%, and 99.83% test accuracy respectively. BHDD is available under CC BY-SA 4.0 at https://github.com/baseresearch/BHDD

We consider the two related problems of detecting if an example is misclassified or out-of-distribution. We present a simple baseline that utilizes probabilities from softmax distributions. Correctly classified examples tend to have greater maximum softmax probabilities than erroneously classified and out-of-distribution examples, allowing for their detection. We assess performance by defining several tasks in computer vision, natural language processing, and automatic speech recognition, showing the effectiveness of this baseline across all. We then show the baseline can sometimes be surpassed, demonstrating the room for future research on these underexplored detection tasks.

Question Answering (QA) returns concise answers or answer lists from natural language text given a context document. Many resources go into curating QA datasets to advance robust models' development. There is a surge of QA datasets for languages like English, however, this is not true for Amharic. Amharic, the official language of Ethiopia, is the second most spoken Semitic language in the world. There is no published or publicly available Amharic QA dataset. Hence, to foster the research in Amharic QA, we present the first Amharic QA (AmQA) dataset. We crowdsourced 2628 question-answer pairs over 378 Wikipedia articles. Additionally, we run an XLMR Large-based baseline model to spark open-domain QA research interest. The best-performing baseline achieves an F-score of 69.58 and 71.74 in reader-retriever QA and reading comprehension settings respectively.

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.

In Reinforcement Learning with Verifiable Rewards (RLVR), constructing a robust advantage baseline is critical for policy gradients, effectively guiding the policy model to reinforce desired behaviors. Recent research has introduced Generalist Value Models (such as V_0), which achieve pre-trained value estimation by explicitly encoding model capabilities in-context, eliminating the need to synchronously update the value model alongside the policy model. In this paper, we propose V_{0.5}, which adaptively fuses the baseline predicted by such value model (acting as a prior) with the empirical mean derived from sparse rollouts. This constructs a robust baseline that balances computational efficiency with extremely low variance. Specifically, we introduce a real-time statistical testing and dynamic budget allocation. This balances the high variance caused by sparse sampling against the systematic bias (or hallucinations) inherent in the value model's prior. By constructing a hypothesis test to evaluate the prior's reliability in real-time, the system dynamically allocates additional rollout budget on demand. This mechanism minimizes the baseline estimator's Mean Squared Error (MSE), guaranteeing stable policy gradients, even under extreme sparsity with a group size of 4. Extensive evaluations across six mathematical reasoning benchmarks demonstrate that V_{0.5} significantly outperforms GRPO and DAPO, achieving faster convergence and over some 10% performance improvement.

We explore the problem of generating minority samples using diffusion models. The minority samples are instances that lie on low-density regions of a data manifold. Generating a sufficient number of such minority instances is important, since they often contain some unique attributes of the data. However, the conventional generation process of the diffusion models mostly yields majority samples (that lie on high-density regions of the manifold) due to their high likelihoods, making themselves ineffective and time-consuming for the minority generating task. In this work, we present a novel framework that can make the generation process of the diffusion models focus on the minority samples. We first highlight that Tweedie's denoising formula yields favorable results for majority samples. The observation motivates us to introduce a metric that describes the uniqueness of a given sample. To address the inherent preference of the diffusion models w.r.t. the majority samples, we further develop minority guidance, a sampling technique that can guide the generation process toward regions with desired likelihood levels. Experiments on benchmark real datasets demonstrate that our minority guidance can greatly improve the capability of generating high-quality minority samples over existing generative samplers. We showcase that the performance benefit of our framework persists even in demanding real-world scenarios such as medical imaging, further underscoring the practical significance of our work. Code is available at https://github.com/soobin-um/minority-guidance.

The 2024 Department of Justice antitrust complaint against RealPage, Inc. named five major residential REITs for coordinating algorithmic rent pricing across hundreds of thousands of apartment units in major US metropolitan areas. This paper studies whether census-tract-level corporate landlord concentration (CLC), measured from SEC EDGAR 10-K property filings geocoded to census tracts, the first such application in the literature, is associated with rent growth 2019-2023, and whether that association is larger in majority-minority neighborhoods. Rent outcomes are measured using the Zillow Observed Rent Index (ZORI). To account for the possibility that corporate landlords preferentially locate in neighborhoods already seeing rent appreciation, all regressions control for a fully novel Algorithmic Housing Burden Index (AHBI), a composite of pre-existing rent burden and market tightness from ACS data. Across 665 census tracts in ten US metropolitan areas, doubling REIT concentration is associated with 2.8 percentage points higher rent growth (p = 0.086, p = 0.030, HC1 robust). This association is significantly stronger in majority-minority tracts. Within the same metro, high-CLC majority-minority tracts are associated with 5.9 percentage points higher rent growth than comparable white tracts (p = 0.039). An XGBoost model predicts 44 percent of out-of-sample rent growth variance, with SHAP analysis independently confirming that CLC's contribution is positive in minority tracts and negative in white tracts. Taken all together, these findings provide the first tract-level evidence consistent with corporate landlord concentration being associated with disproportionately higher rent growth in communities of color.

As large language models increase in capability, researchers have started to conduct surveys of all kinds on these models with varying scientific motivations. In this work, we examine what we can learn from a model's survey responses on the basis of the well-established American Community Survey (ACS) by the U.S. Census Bureau. Evaluating more than a dozen different models, varying in size from a few hundred million to ten billion parameters, hundreds of thousands of times each on questions from the ACS, we systematically establish two dominant patterns. First, smaller models have a significant position and labeling bias, for example, towards survey responses labeled with the letter "A". This A-bias diminishes, albeit slowly, as model size increases. Second, when adjusting for this labeling bias through randomized answer ordering, models still do not trend toward US population statistics or those of any cognizable population. Rather, models across the board trend toward uniformly random aggregate statistics over survey responses. This pattern is robust to various different ways of prompting the model, including what is the de-facto standard. Our findings demonstrate that aggregate statistics of a language model's survey responses lack the signals found in human populations. This absence of statistical signal cautions about the use of survey responses from large language models at present time.

State-of-the-art generative text-to-image models are known to exhibit social biases and over-represent certain groups like people of perceived lighter skin tones and men in their outcomes. In this work, we propose a method to mitigate such biases and ensure that the outcomes are fair across different groups of people. We do this by finetuning text-to-image models on synthetic data that varies in perceived skin tones and genders constructed from diverse text prompts. These text prompts are constructed from multiplicative combinations of ethnicities, genders, professions, age groups, and so on, resulting in diverse synthetic data. Our diversity finetuned (DFT) model improves the group fairness metric by 150% for perceived skin tone and 97.7% for perceived gender. Compared to baselines, DFT models generate more people with perceived darker skin tone and more women. To foster open research, we will release all text prompts and code to generate training images.

Large participatory biomedical studies, studies that recruit individuals to join a dataset, are gaining popularity and investment, especially for analysis by modern AI methods. Because they purposively recruit participants, these studies are uniquely able to address a lack of historical representation, an issue that has affected many biomedical datasets. In this work, we define representativeness as the similarity to a target population distribution of a set of attributes and our goal is to mirror the U.S. population across distributions of age, gender, race, and ethnicity. Many participatory studies recruit at several institutions, so we introduce a computational approach to adaptively allocate recruitment resources among sites to improve representativeness. In simulated recruitment of 10,000-participant cohorts from medical centers in the STAR Clinical Research Network, we show that our approach yields a more representative cohort than existing baselines. Thus, we highlight the value of computational modeling in guiding recruitment efforts.

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.

Pir\'a is a reading comprehension dataset focused on the ocean, the Brazilian coast, and climate change, built from a collection of scientific abstracts and reports on these topics. This dataset represents a versatile language resource, particularly useful for testing the ability of current machine learning models to acquire expert scientific knowledge. Despite its potential, a detailed set of baselines has not yet been developed for Pir\'a. By creating these baselines, researchers can more easily utilize Pir\'a as a resource for testing machine learning models across a wide range of question answering tasks. In this paper, we define six benchmarks over the Pir\'a dataset, covering closed generative question answering, machine reading comprehension, information retrieval, open question answering, answer triggering, and multiple choice question answering. As part of this effort, we have also produced a curated version of the original dataset, where we fixed a number of grammar issues, repetitions, and other shortcomings. Furthermore, the dataset has been extended in several new directions, so as to face the aforementioned benchmarks: translation of supporting texts from English into Portuguese, classification labels for answerability, automatic paraphrases of questions and answers, and multiple choice candidates. The results described in this paper provide several points of reference for researchers interested in exploring the challenges provided by the Pir\'a dataset.

Knowledge tracing (KT) is the problem of predicting students' future performance based on their historical interactions with intelligent tutoring systems. Recently, many works present lots of special methods for applying deep neural networks to KT from different perspectives like model architecture, adversarial augmentation and etc., which make the overall algorithm and system become more and more complex. Furthermore, due to the lack of standardized evaluation protocol liu2022pykt, there is no widely agreed KT baselines and published experimental comparisons become inconsistent and self-contradictory, i.e., the reported AUC scores of DKT on ASSISTments2009 range from 0.721 to 0.821 minn2018deep,yeung2018addressing. Therefore, in this paper, we provide a strong but simple baseline method to deal with the KT task named simpleKT. Inspired by the Rasch model in psychometrics, we explicitly model question-specific variations to capture the individual differences among questions covering the same set of knowledge components that are a generalization of terms of concepts or skills needed for learners to accomplish steps in a task or a problem. Furthermore, instead of using sophisticated representations to capture student forgetting behaviors, we use the ordinary dot-product attention function to extract the time-aware information embedded in the student learning interactions. Extensive experiments show that such a simple baseline is able to always rank top 3 in terms of AUC scores and achieve 57 wins, 3 ties and 16 loss against 12 DLKT baseline methods on 7 public datasets of different domains. We believe this work serves as a strong baseline for future KT research. Code is available at https://github.com/pykt-team/pykt-toolkitWe merged our model to the \textsc{pyKT benchmark at https://pykt.org/.}.

Large language model (LLM) agents increasingly take actions (screening applicants, recommending credit, triaging patients), yet fairness for LLMs is still measured by grading answers. We introduce AgentFairBench, a cheap, reproducible, multi-domain benchmark for demographic disparity in the actions of LLM agents. Grounded in a companion framework, the Bias Conduction Framework (BCF, restated here), it spans three regulator-anchored domains: hiring, lending, and medical triage. Synthetic, demographic-neutral profiles are evaluated in counterfactual matched sets that vary only a name-coded race x gender signal (in the Bertrand Mullainathan tradition), under four agent scaffolds of increasing agency (direct, chain-of-thought, multi-agent deliberation, tool-augmented). A NumPy-only harness computes counterfactual flip rate, mean absolute score difference (MASD), action-rate disparity, and tool-invocation disparity, with bootstrap confidence intervals, paired tests, and false-discovery-rate control, for single-digit dollars per model. A live leaderboard with a held-out private split and a contamination canary admits external models by submission. Our pilot (864 decisions plus a test-retest replication) carries a methodological lesson: comparing a six-group score spread against a two-run noise difference overstates disparity by ~ 2.4X through statistic arity alone. Against an arity matched noise floor and an omnibus group test, claude haiku 4 5 shows no demographic effect above sampling noise (0 of 120 pairwise and 0 of 9 omnibus contrasts survive correction); a planted-bias test confirms the instrument detects disparity when present. The contribution is a sound, sensitive, adoption-ready instrument, the arity matched null methodology, and open artifacts to scale it. Code, data, and harness are released under open licenses, with an anonymized review artifact.

Claims that language models homogenise are usually measured against human judgements collected for the study, which makes the human side an artifact of the design: a crowdworker given the model's instruction is running the model's prompt. We measure convergence against a human reference nobody built for the purpose -- 2,523 reader mark sets across 120 web documents, produced by people highlighting for their own reasons on a platform where the overlay of others' marks is off by default. Agreement is the overlap between two size-matched sentence sets minus the overlap expected when each is resampled within its own depth-and-length bands. The null's calibration is demonstrated, not asserted: every pair involving a random baseline lands within 0.006 of zero. On the median document each party names 14 sentences of 70; two readers share 4.1 and two models 8.7. Across 18 model arms spanning 11 vendors, 3 countries and both weight regimes, the median of 153 model pairs is +0.093 against a human yardstick of +0.040, and 99 sit entirely above the human interval. Two frontier models from rival labs reach +0.203, twice what GPT-4o agrees with itself on a second call. The effect is not determinism, prompt wording, procedure, vendor or routing, and it is graded: the smallest models agree at the human level. No model agrees with readers detectably more than a reader does, and at equal depth and length no surface feature separates their choices. The multiples are procedure-dependent and the ordering is not: models are cut to their sharpest set while a reader's is a random draw from what they marked, and blunting the models alike halves the gap without closing it. Tested out of sample on four models released after this analysis, against predictions fixed beforehand, none clears the human interval. A population simulated from several models is not several populations.

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.

The task of mixture proportion estimation (MPE) is to estimate the weight of a component distribution in a mixture, given observations from both the component and mixture. Previous work on MPE adopts the irreducibility assumption, which ensures identifiablity of the mixture proportion. In this paper, we propose a more general sufficient condition that accommodates several settings of interest where irreducibility does not hold. We further present a resampling-based meta-algorithm that takes any existing MPE algorithm designed to work under irreducibility and adapts it to work under our more general condition. Our approach empirically exhibits improved estimation performance relative to baseline methods and to a recently proposed regrouping-based algorithm.

Fashion attribute extraction is evaluated inconsistently: results are reported as single aggregate numbers across image types that pose different problems, fields that are not visible in an image are scored as ordinary negatives, and the effect of vocabulary mismatch between datasets is acknowledged but not measured. We release the MODA General Attribute Suite, a four-track benchmark that keeps these problems separate by construction. Each track (localized garment crops, catalogue product images, applicability-aware full-body photographs, and product text) carries its own frozen test set, input contract, metric, and leakage unit, and the tracks are never averaged. The protocol requires label-blind prediction, SHA-256 commitment of prediction files before any label is opened, a fail-closed scorer, and a 10,000-sample paired cluster bootstrap at each track's natural leakage unit; promotion requires a positive interval on every track rather than a favourable mean. We release the scorers, the split builders, our own prediction files with their hashes including the runs we lose, and the MODA_NER(V) model checkpoints for three of the four tracks. The text model is not distributed; its benchmark is. We report baseline results, six interventions that did not improve them, and the limitations of the evidence.

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%).

Group-relative policy optimization methods train language models by generating multiple rollouts per prompt and normalizing rewards with a shared mean reward baseline. In resource-constrained settings where the rollout budget is small, accuracy often degrades. We find that noise in the shared baseline induces advantage sign flips, where some rollouts receive an incorrect advantage sign, and the update direction is reversed. To address this, we propose Median-Centered Group Relative Policy Optimization (MC-GRPO), a simple and effective solution for small-rollout training. Our main idea is to replace the mean baseline with a median baseline: the median is far less sensitive to outlier rewards than the mean, mitigating the sign flips under small rollout size (G). We generate one additional rollout for median reference (G+1), and compute advantages by using the group median. With an odd-sized group, exactly one completion is the median and receives zero advantage, we exclude this pivot rollout from backpropagation so the number of gradient-contributing samples per prompt remains G, preserving the core update cost of standard G-rollout training. Across various GRPO-family methods and a wide range of models and scales, this median-centered training consistently improves stability and final accuracy in the low-rollout regime, reducing the gap between G=2 and G=8 to within 1%. Code is available at https://github.com/lotusroot-kim/MC-GRPO

Whether a language model has improved itself is increasingly judged not by mean accuracy but by which individual problems it gains and loses. Tracking these transitions means differencing two noisy estimates, leaving them vulnerable to measurement artifacts. Auditing three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control pushed through the identical pipeline, we identify seven measurement failures, each of which inverts a reported finding when its control is absent. Several are standard practice. A ledger built on a single greedy decode manufactures capability changes on an untrained model, largely an artifact of inference batching; the expansion statistic separating acquisition from sharpening assigns that same model a rate of 0.280. The natural threshold repair does not survive replication: estimated across the frozen comparisons such a design already contains, its null stays non-zero. We replace it with a per-problem exact test against a pooled baseline under false-discovery-rate control, which detects nothing on any held-out replicate and is unchanged under the multiple-testing rule, error rate and pool size. Applied to a ladder of arms matched in stream, volume and evaluation, the audit finds that external distillation improves problems the base model rarely reaches while three forms of self-training do not; a regression rejects this asymmetry as a by-product of distillation's larger overall gain (p < 10^{-8}). On the far smaller set of problems the base model never reaches, the evidence is inconclusive, while self-training corrupts problems solved at baseline at rates well above the measured floor. Transition-level auditing therefore requires a separately measured null for every statistic it reports: nulls that cost no new experiments, built from baseline replicates a multi-arm study already owns, though not from as few as most possess.

When answering a question, people often draw upon their rich world knowledge in addition to the particular context. Recent work has focused primarily on answering questions given some relevant document or context, and required very little general background. To investigate question answering with prior knowledge, we present CommonsenseQA: a challenging new dataset for commonsense question answering. To capture common sense beyond associations, we extract from ConceptNet (Speer et al., 2017) multiple target concepts that have the same semantic relation to a single source concept. Crowd-workers are asked to author multiple-choice questions that mention the source concept and discriminate in turn between each of the target concepts. This encourages workers to create questions with complex semantics that often require prior knowledge. We create 12,247 questions through this procedure and demonstrate the difficulty of our task with a large number of strong baselines. Our best baseline is based on BERT-large (Devlin et al., 2018) and obtains 56% accuracy, well below human performance, which is 89%.

Large language models (LLMs) are increasingly used as judges for subjective tasks, where annotators disagree and the relevant question is not only how accurate a judge is, but whose judgments it reproduces. Sociodemographic prompting conditions the judge on an annotator's demographic profile to align its judgments with the corresponding group's. We test whether this alignment emerges distributionally, comparing the predicted label distributions of 23 open-weight LLMs on three subjective tasks against those of real annotator groups, under three conditions: no demographic information, single-attribute profiles, and intersectional profiles over gender, age, race, and education. Three findings emerge. First, a judge prompted with no demographics is not perspective-neutral: models best reproduce the judgments of White, college-educated annotators. Second, demographic conditioning is asymmetric: it moves the judge toward majority groups and away from minority groups, most strongly on offensiveness, where intersectional profiles amplify the harm. Third, by comparing base and instruct models we identify instruction-tuning as a possible source of the asymmetry. Demographic conditioning should therefore be used with caution to estimate group judgments: conditioning moves predictions away from the reference distributions of the minority groups the method is often invoked to serve.

Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. We also demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.

In-context learning is commonly interpreted as a form of conditional inference, in which the prompt specifies a context and the model's output is treated as an estimate of the corresponding conditional distribution. If this interpretation holds, then LLM estimates should satisfy basic probabilistic identities. In particular, the law of total probability asserts that prior-weighted conditional distributions aggregate into population-level marginals over any valid partition of the population. In this work, we investigate to what extent LLM estimates adhere to this self-consistency principle. We use binary trees as an evaluation scaffold to recursively partition a population into increasingly fine-grained subpopulations. We then prompt LLMs with verbalized subpopulation descriptions in context, aggregate the resulting estimates back into population-level estimates, and compare them across partitions of varying granularity. Applying this protocol across problem domains and state-of-the-art frontier models, we show widespread violations of basic consistency properties. An in-depth study of persona prompting reveals a pattern we call the macro fallacy: estimates reconstructed from more fine-grained subpopulation responses are often better aligned with human reference data than direct population-level estimates. This effect persists across variations in tree structure and estimation task, and can be partially recovered through implicit prompting. Together, these findings suggest that models possess relevant subpopulation knowledge but do not reliably propagate it into aggregate estimates. This gap establishes statistical self-consistency as an unsaturated, reference-free criterion for evaluating LLMs.

Large pre-trained language models have shown promise for few-shot learning, completing text-based tasks given only a few task-specific examples. Will models soon solve classification tasks that have so far been reserved for human research assistants? Existing benchmarks are not designed to measure progress in applied settings, and so don't directly answer this question. The RAFT benchmark (Real-world Annotated Few-shot Tasks) focuses on naturally occurring tasks and uses an evaluation setup that mirrors deployment. Baseline evaluations on RAFT reveal areas current techniques struggle with: reasoning over long texts and tasks with many classes. Human baselines show that some classification tasks are difficult for non-expert humans, reflecting that real-world value sometimes depends on domain expertise. Yet even non-expert human baseline F1 scores exceed GPT-3 by an average of 0.11. The RAFT datasets and leaderboard will track which model improvements translate into real-world benefits at https://raft.elicit.org .

The language identification task is a crucial fundamental step in NLP. Often it serves as a pre-processing step for widely used NLP applications such as multilingual machine translation, information retrieval, question and answering, and text summarization. The core challenge of language identification lies in distinguishing languages in noisy, short, and code-mixed environments. This becomes even harder in case of diverse Indian languages that exhibit lexical and phonetic similarities, but have distinct differences. Many Indian languages share the same script making the task even more challenging. In this paper, we release a dataset of 230K sentences consisting of English and all 22 official Indian languages labeled with their language identifiers where data in most languages are newly created. We also develop and release robust baseline models using state-of-the-art approaches in machine learning and deep learning that can aid the research in this field. Our baseline models are comparable to the state-of-the-art models for the language identification task.

Data-centric curation pipelines frequently rely on model confidence scores to flag and filter noisy or mislabeled training instances. Evaluating this filtering convention on a large-scale consumer lending sample (LendingClub, N = 1,344,936) uncovers an underlying demographic asymmetry: high-income defaulters are disproportionately classified as label noise relative to low-income defaulters (Cramer's V approximately 0.03-0.07). Re-examining this behavior through the lens of equal opportunity [Hardt et al., 2016] reveals a far more severe discrepancy: a 16.86 percentage point gap in true positive rate (recall) between high- and low-income borrowers who ultimately defaulted. Implementing a sequential feature-blinding methodology allows us to isolate the drivers of this disparity across three distinct mechanisms: (1) direct reliance on self-reported applicant income; (2) algorithmic absorption of upstream institutional bias encoded within origination interest rates; and (3) a residual disparity (3.55 percentage points in cross-validation; 2.56 percentage points on a held-out test partition, Z = -4.04, p < 0.0001) that remains even after purging both income and interest rates from the model. Out-of-sample signed SHAP valuations demonstrate that this residual gap is maintained by structural proxies, most notably loan amount and home ownership status. These empirical findings show that simply blinding an algorithm to sensitive attributes fails to ensure fairness when institutional pricing decisions and behavioral proxy variables collectively reconstruct the omitted signals. We outline the practical implications of these findings for auditing data-centric AI workflows within regulated financial institutions.

We present PopResume, a population-representative resume dataset for causal fairness auditing of LLM- and VLM-based resume screening systems. Unlike existing benchmarks that rely on manually injected demographic information and outcome-level disparities, PopResume is grounded in population statistics and preserves natural attribute relationships, enabling path-specific effect (PSE)-based fairness evaluation. We decompose the effect of a protected attribute on resume scores into two paths: the business necessity path, mediated by job-relevant qualifications, and the redlining path, mediated by demographic proxies. This distinction allows auditors to separate legally permissible from impermissible sources of disparity. Evaluating four LLMs and four VLMs on PopResume's 60.8K resumes across five occupations, we identify five representative discrimination patterns that aggregate metrics fail to capture. Our results demonstrate that PSE-based evaluation reveals fairness issues masked by outcome-level measures, underscoring the need for causally-grounded auditing frameworks in AI-assisted hiring.

This paper uses the Korean Baseball Organization's adoption of the Automated Ball-Strike (ABS) system to audit long-standing claims about contextual bias in human ball-strike calls. Using pitch-level KBO data from 2021 through the available portion of the 2026 season, we model called-strike probability for taken pitches near the strike-zone boundary, with 2022-2023 as the primary human-umpire baseline and ABS seasons (2024 and onward) as a diagnostic benchmark. The strongest evidence concerns count pressure. Relative to 0--0 counts, human umpires called substantially fewer strikes in two-strike counts and more strikes in hitter-ahead three-ball counts. Specifically, in the main 0.25-ft boundary band, 0--2 was associated with a -17.17 percentage-point effect and 3--0 with a +6.61 percentage-point effect. Under ABS, the corresponding effects were close to zero and did not survive false-discovery-rate correction. Game progression shows a smaller but coherent pattern as human calls were less strike-prone in early innings and more strike-prone in innings 7--9+, especially in late-close situations, while complete ABS seasons were essentially flat. Other suspected biases are weaker or more localized. Salary-based reputation proxies provide suggestive but proxy-sensitive evidence, and catcher identity shows human-period residual heterogeneity that disappears under ABS. Home-context evidence is mostly null at the umpire level, with one FDR-significant human-period exception and an exploratory umpire-team gap best treated as an audit lead. Overall, the results do not show that human umpires were biased everywhere. Instead, they map where the human strike zone was most context-sensitive, where evidence was weaker, and where common suspicions received little support.

We revisit the performance of the classic gradual magnitude pruning (GMP) baseline for large language models, focusing on the classic BERT benchmark on various popular tasks. Despite existing evidence in the literature that GMP performs poorly, we show that a simple and general variant, which we call GMP*, can match and sometimes outperform more complex state-of-the-art methods. Our results provide a simple yet strong baseline for future work, highlight the importance of parameter tuning for baselines, and even improve the performance of the state-of-the-art second-order pruning method in this setting.

Demographic synthetic survey panels are often validated by matching aggregate answers to published surveys. We test what that certificate establishes across six multiselect batteries from four survey organisations in three countries. The headline analysis is restricted to three instruments whose synthetic cohort and human target share the stated population frame; three other batteries remain sensitivity analyses. The response contract dominates measured fidelity. In the aligned instruments, committed sets leave 66 of 128 model-battery option slots empty in panels of up to 500 respondents, versus 0 of 128 under per-option probability elicitation. Across eight uncapped model-instrument comparisons, probabilities reduce option-marginal MAE by 4.53 to 7.30 points. The capped instrument reverses on two models until the vectors are projected onto its stated maximum. These are measurement effects: human targets are realised check-all responses, whereas the vectors are latent inclusion propensities. Published marginal agreement also fails to discriminate respondent simulation from direct population estimation. On nine aligned model-battery pairs, a no-persona population-prevalence query averages 6.27 MAE versus 12.39 for committed panels and wins all nine comparisons. Constraint-aware probability vectors average 5.34 and beat the query on four of nine, so the baseline challenges the validation criterion rather than proving direct estimation uniformly best. On three unpublished demographic cells, neither approach beats reciting the national distribution. Population-marginal agreement is therefore evidence about an elicitation contract and an estimand obtainable without simulated respondents, not evidence of individual simulation.

This paper introduces a new open-source speech corpus named "speechocean762" designed for pronunciation assessment use, consisting of 5000 English utterances from 250 non-native speakers, where half of the speakers are children. Five experts annotated each of the utterances at sentence-level, word-level and phoneme-level. A baseline system is released in open source to illustrate the phoneme-level pronunciation assessment workflow on this corpus. This corpus is allowed to be used freely for commercial and non-commercial purposes. It is available for free download from OpenSLR, and the corresponding baseline system is published in the Kaldi speech recognition toolkit.

Objective: To propose and retrospectively validate an integrated framework addressing three barriers to clinical translation of readmission prediction: lack of explainability, absence of deployment reliability infrastructure, and inadequate demographic fairness evaluation. Materials and Methods: We constructed a cohort of 415231 adult admissions from the MIMIC-IV database (30-day readmission prevalence 18.0%), split 70/15/15. Logistic regression, XGBoost, and LightGBM models were trained on 26 features. SHAP provided per-patient explanations. Fairness was evaluated across 16 subgroups using AUC-ROC, false negative rate (FNR), and positive predictive value (PPV). Calibration was assessed using Brier scores and calibration curves. Results: XGBoost achieved AUC-ROC 0.696 (95% CI 0.691-0.701), outperforming or matching the LACE baseline (AUC 0.60-0.68). LightGBM achieved best calibration (Brier 0.146). Prior admissions were the dominant predictor. All subgroups met equity thresholds (delta AUC <= 0.05, delta FNR <= 0.10). Conclusion: This framework delivers competitive performance, clinically actionable explanations, and strong demographic equity. Code is publicly available at https://github.com/Tomisin92/readmission-prediction.

Self-consistency (SC) is a widely used test-time inference technique for improving performance in chain-of-thought reasoning. It involves generating multiple responses, or samples from a large language model (LLM) and selecting the most frequent answer. This procedure can naturally be viewed as a majority vote or empirical mode estimation. Despite its effectiveness, SC is prohibitively expensive at scale when naively applied to datasets, and it lacks a unified theoretical treatment of sample efficiency and scaling behavior. In this paper, we provide the first comprehensive analysis of SC's scaling behavior and its variants, drawing on mode estimation and voting theory. We derive and empirically validate power law scaling for self-consistency across datasets, and analyze the sample efficiency for fixed-allocation and dynamic-allocation sampling schemes. From these insights, we introduce Blend-ASC, a novel variant of self-consistency that dynamically allocates samples to questions during inference, achieving state-of-the-art sample efficiency. Our approach uses 6.8x fewer samples than vanilla SC on average, outperforming both fixed- and dynamic-allocation SC baselines, thereby demonstrating the superiority of our approach in terms of efficiency. In contrast to existing variants, Blend-ASC is hyperparameter-free and can fit an arbitrary sample budget, ensuring it can be easily applied to any self-consistency application.

Individuals turn to large language models (LLMs) for guidance across a wide range of economic tasks, from comparing loan options and planning savings to deciding what raise to ask for or how much to charge for their services. LLMs are known to reproduce social biases, and biased economic guidance may influence what users believe they are worth, what they ask for, and what they ultimately accept. This risk is especially salient in India, where economic outcomes are shaped by demographic categories such as caste and urban-rural location. Existing LLM bias benchmarks, however, are largely designed around Western demographic categories and therefore miss key axes of economic disparity in the Indian context. We introduce RupeeBias, a benchmark for auditing demographic bias in LLM-generated economic guidance across Indian economic settings. RupeeBias consists of 39,150 prompts spanning four use cases: salary estimation, salary increment estimation, counter-offer recommendation, and service pricing recommendation. The benchmark follows a single-attribute counterfactual design, holding the description of the user's qualifications, experience, or service offering fixed while varying one demographic identifier at a time. RupeeBias covers 87 India-specific demographic identifiers across six axes: caste, religion, regional identity, gender, disability, and urban-rural location, with all prompts constructed in both English and Hinglish. We evaluate nine LLMs on RupeeBias and find systematic demographic disparities across all six axes. For otherwise identical prompts that differ only in demographic identifier, LLM-generated economic outputs differ by 20.2% on average. We publicly release RupeeBias to support future research on demographic bias in LLM-generated economic guidance across India-specific demographic and economic contexts.

In Machine Translation (MT) evaluation, metric performance is assessed based on agreement with human judgments. In recent years, automatic metrics have demonstrated increasingly high levels of agreement with humans. To gain a clearer understanding of metric performance and establish an upper bound, we incorporate human baselines in the MT meta-evaluation, that is, the assessment of MT metrics' capabilities. Our results show that human annotators are not consistently superior to automatic metrics, with state-of-the-art metrics often ranking on par with or higher than human baselines. Despite these findings suggesting human parity, we discuss several reasons for caution. Finally, we explore the broader implications of our results for the research field, asking: Can we still reliably measure improvements in MT evaluation? With this work, we aim to shed light on the limits of our ability to measure progress in the field, fostering discussion on an issue that we believe is crucial to the entire MT evaluation community.

DARTS search space (DSS) has become a canonical benchmark for NAS whereas some emerging works pointed out the issue of narrow accuracy range and claimed it would hurt the method ranking. We observe some recent studies already suffer from this issue that overshadows the meaning of scores. In this work, we first propose and orchestrate a suite of improvements to frame a larger and harder DSS, termed LHD, while retaining high efficiency in search. We step forward to renovate a LHD-based new benchmark, taking care of both discernibility and accessibility. Specifically, we re-implement twelve baselines and evaluate them across twelve conditions by combining two underexpolored influential factors: transductive robustness and discretization policy, to reasonably construct a benchmark upon multi-condition evaluation. Considering that the tabular benchmarks are always insufficient to adequately evaluate the methods of neural architecture search (NAS), our work can serve as a crucial basis for the future progress of NAS. https://github.com/chaoji90/LHD

Recent autonomous penetration testing papers report high benchmark scores while adding multi-component security harnesses around frontier LLMs. Because these systems often change both architecture and backbone model, it is difficult to tell how much performance comes from the harness rather than from the underlying model. This paper presents a controlled study on the 104-task XBOW benchmark using default coding CLI agents as plain-agent baselines. We first run Codex, OpenCode, and Pi with the same GPT-5 model, budget, target interface, and scoring rule. This phase identifies the strongest same-model baseline and tests whether security-specific prompt variants improve its observed score. We then compare the default Codex scaffold with published MAPTA and PentestGPT V2 results under the closest available model matches. Finally, we repeat the plain-agent experiment with GPT-5.2 and GPT-5.5 to measure model scaling inside the same scaffold. The results show a mixed but practical picture. Specialised harnesses can add measurable benchmark lift and may improve cost efficiency, but plain coding agents already solve a large share of the benchmark; repeated plain-agent runs can match or exceed some published architecture scores in union coverage, and newer models substantially improve the same scaffold. Future evaluations should report model-matched plain-agent baselines before attributing benchmark gains to architecture design alone.

This study developed a new Bangla abstractive summarization dataset to generate concise summaries of Bangla articles from diverse sources. Most existing studies in this field have concentrated on news articles, where journalists usually follow a fixed writing style. While such approaches are effective in limited contexts, they often fail to adapt to the varied nature of real-world Bangla texts. In today's digital era, a massive amount of Bangla content is continuously produced across blogs, newspapers, and social media. This creates a pressing need for summarization systems that can reduce information overload and help readers understand content more quickly. To address this challenge, we developed a dataset of over 54,000 Bangla articles and summaries collected from multiple sources, including blogs such as Cinegolpo and newspapers such as Samakal and The Business Standard. Unlike single-domain resources, our dataset spans multiple domains and writing styles. It offers greater adaptability and practical relevance. To establish strong baselines, we trained and evaluated this dataset using several deep learning and transfer learning models, including LSTM, BanglaT5-small, and MTS-small. The results highlight its potential as a benchmark for future research in Bangla natural language processing. This dataset provides a solid foundation for building robust summarization systems and helps expand NLP resources for low-resource languages.

Prior research has established that instruction-tuned large language models exhibit left-of-center political bias, measured exclusively through abstract political questionnaires. We show that this finding does not generalize to concrete policy decisions. We introduce a dual-instrument methodology grounded in Swiss democratic reality. The Smartvote questionnaire (75 abstract policy questions) is administered to 66 LLMs from 27 model families and compared to 184 elected members of the Swiss National Council, replicating the established leftward convergence (Cohen's d = 3.64, p = 0.0002). Then, novel to this work, 9 flagship LLMs are confronted with 48 real federal referenda (Volksabstimmungen) in four national languages (German, French, Italian, Romansh) under three information conditions, comparing votes to actual outcomes and party recommendations (Parolen). Three findings challenge the prevailing narrative. (1) Abstract questionnaires do not predict concrete behavior: the left-to-right agreement gradient on Smartvote shifts from left-peaked to center-peaked on Volksabstimmungen, where models align most with centrist Die Mitte and FDP rather than leftist SP and Gruene (Wilcoxon p = 0.008). (2) For some models, the language of a political question changes the answer more than the political content does: cross-linguistic consistency ranges from 50% (Mistral) to 98% (GPT-5.4). (3) Two models exhibit systematic change-aversion rather than political bias, voting Nein on 83-94% of referenda regardless of direction (binomial p < 0.0001). What prior work measured as "leftward bias" may not generalize beyond abstract instruments. On concrete policy decisions, LLMs behave less like coalition partners of the left and more like cautious civil servants: centrist, status-quo-favoring, and inconsistent across languages.

We measure the emotion geometry of six LLMs spanning Google's Gemma 2 family. Using an extraction pipeline established by Sofroniew et al. (2026) on Claude Sonnet 4.5 and replicated by Hollowell (2026) on Gemma 2 27B base, we compare emotion geometry across base and instruction-tuned variants of Gemma 2 2B, 9B, and 27B. First, all six model cells exhibit the core emotion geometry. A linear map fit between two cells places held-out emotions nearest their own vectors in the target model. Across all fifteen model pairs, fit in one direction each, this failed exactly once, and the median rank was 1 in every pair. Linear maps refit on cleaned embedding-table vectors recover every held-out emotion at rank 1 in all thirty ordered pairs. Second, scale spreads the geometry's variance across more directions without weakening its tie to valence. The leading component's share of variance falls from 31.6 percent (2B) to 29.8 (9B) to 26.7 (27B) in the base models and from 31.9 (2B) to 29.2 (9B) to 27.3 (27B) in the tuned ones, while its correlation with human valence ratings stays between .71 and .78 in all six models. Third, we project each emotion vector onto the output vocabulary and ask at what depth the vector's own emotion word enters the top 10,000 of the 256,000 tokens. In the 2B and 9B models the median emotion word is there from layer 0, while in the 27B models it begins near chance, stays outside the top 10,000 for half the network, and enters at layer 23 in one step. Fourth, tuning leaves every registered analysis-layer measure within its floor. At every size, valence correlation rises and arousal correlation falls by amounts within their floors, the latter against prediction. PC2's variance share also falls, with a pooled decrease significant after correction. Mid-late layer stability exceeds its floor and traces to base 27B's final layer. Code, data, and audit reports are released.

Vision-language models (VLMs) are increasingly used to make decisions from visual inputs. We introduce FAIRLENS, a benchmark and evaluation framework for measuring both the fairness and the validity of VLM responses in three high-stakes domains: hiring, legal, and healthcare. FAIRLENS pairs real face images spanning gender, race, and age groups with closed- and open-ended questions, giving more than 100K image-question pairs per model, and evaluates responses from four complementary views: demographic parity over adverse outcome rates, soundness, demographic association over unsupported roles and statuses, and bias in free-text generation. Soundness is the central validity criterion: a response is sound when it follows the evidence stated in the question and abstains when the image cannot support an answer. Evaluating eight VLMs, we find that the primary failure is unwarranted inference rather than unequal treatment. Models routinely infer qualifications, threat, illness, or professional role from a face instead of abstaining, and the weakest model does so on 99% of the questions its input cannot answer. These failures are most severe in legal and healthcare, where recognizing insufficient evidence matters most, and disparity metrics alone would miss them: parity gaps are small in absolute terms, yet when baseline adverse rates are low the same gap means one demographic group receives adverse labels several times as often as another, and a small gap can equally reflect a model that treats every group unsafely. Bias in free-text responses is only loosely coupled to multiple-choice accuracy, so correct structured answers do not imply safe generation. FAIRLENS shows that fair high-stakes VLM behavior requires similar treatment across groups and refusal to infer high-stakes attributes from appearance, and its question suite transfers to any face corpus with demographic annotations.

Large language models (LLMs) are increasingly embedded in educational, professional, and public infrastructure, yet widely used platforms expose "English (US)" as a primary English setting despite the global diversity of English. We ask: How does "English (US)" become the default? We study this question as structural bias, examining how geopolitical histories of data curation, digital dominance, and linguistic standardization intersect with the LLM development pipeline. Using British English as a controlled reference, we construct a curated resource of 1,813 matched American English (AmE)--British English (BrE) variants and introduce DiAlign, a dynamic, training-free method for estimating regional alignment from distributional evidence. We triangulate the AmE preference across data exposure --> representation --> generation, jointly examining pretraining and post-training data, tokenizer behavior and provenance, model prediction cost, and generated language across developer countries, prompt conditions, domains and sources, linguistic categories, and registers. AmE is consistently favored across all six audited pretraining corpora and 21 post-training datasets, is generally represented more compactly by tokenizers, and receives lower prediction cost. It also remains the dominant generation default under neutral English prompting; British-English prompting shifts this preference toward BrE but does not consistently eliminate the AmE default. To our knowledge, this is the first rigorous pipeline-wide study of structural bias across major phases of LLM development. Our findings show that contemporary LLMs privilege AmE as the de facto norm, raising concerns about linguistic homogenization, epistemic injustice, and inequity in global AI deployment, while providing a rigorous basis for targeted component-level intervention.

Majority voting over multiple LLM attempts improves mathematical reasoning, but correlated errors limit the effective sample size. A natural fix is to assign different reasoning strategies to different voters. The approach, Diverse Prompt Mixer, is tested on the AIMO 3 competition: 3 models, 23+ experiments, 50 IMO-level problems, one H100 80 GB, 5-hour limit. Every prompt-level intervention fails. High-temperature sampling already decorrelates errors; weaker strategies reduce accuracy more than they reduce correlation. Across an 8-point capability gap at equal N=8 and every optimization tested, model capability dominates. The gap between the best majority-vote score (42/50) and pass@20 (~45.5) is selection loss, not prompt loss. A verifier-based selector could close it. Prompt engineering cannot.

Urban assessments often compress diverse needs into single scores, which can obscure minority perspectives. We present a community-centered study in Montreal (n=35; wheelchair users, seniors, LGBTQIA2+ residents, and immigrants). Participants rated 20 streets (accessibility, inclusivity, aesthetics, practicality) and ranked 7 images on 12 interview-elicited criteria. Disagreement patterns were systematic in our sample: wheelchair users diverged most on accessibility and practicality; LGBTQIA2+ participants emphasized inclusion and liveliness; seniors prioritized security. Group discussion reduced information gaps but not value conflicts; ratings conveyed intensity, while rankings forced trade-offs. We then formalize negotiative alignment, a transparent, budget-aware bargaining procedure, and pilot it with role-played stakeholder agents plus a neutral mediator. Relative to the best base design under the same public rubric, the negotiated package increased total utility (21.10 to 24.55), raised the worst-group utility (3.20 to 3.90), improved twentieth percentile satisfaction (0.86 to 1.00; min-max normalized within the scenario), and reduced inequality (Gini 0.036 to 0.025). Treating disagreement as signal and reporting worst-group outcomes alongside totals may help planners and AI practitioners surface trade-offs and preserve minority priorities while maintaining efficiency.

The Languini Kitchen serves as both a research collective and codebase designed to empower researchers with limited computational resources to contribute meaningfully to the field of language modelling. We introduce an experimental protocol that enables model comparisons based on equivalent compute, measured in accelerator hours. The number of tokens on which a model is trained is defined by the model's throughput and the chosen compute class. Notably, this approach avoids constraints on critical hyperparameters which affect total parameters or floating-point operations. For evaluation, we pre-process an existing large, diverse, and high-quality dataset of books that surpasses existing academic benchmarks in quality, diversity, and document length. On it, we compare methods based on their empirical scaling trends which are estimated through experiments at various levels of compute. This work also provides two baseline models: a feed-forward model derived from the GPT-2 architecture and a recurrent model in the form of a novel LSTM with ten-fold throughput. While the GPT baseline achieves better perplexity throughout all our levels of compute, our LSTM baseline exhibits a predictable and more favourable scaling law. This is due to the improved throughput and the need for fewer training tokens to achieve the same decrease in test perplexity. Extrapolating the scaling laws leads of both models results in an intersection at roughly 50,000 accelerator hours. We hope this work can serve as the foundation for meaningful and reproducible language modelling research.

LLM inference often generates a batch of candidates for a prompt and selects one via strategies like majority voting or Best-of- N (BoN). For difficult tasks, this single-shot selection often underperforms. Consequently, evaluations commonly report Pass@k: the agent may submit up to k responses, and only the best of them is used when computing regret. Motivated by this, we study inference scaling in the more general Pass@k inference setting, and prove that neither majority voting nor BoN exhibits the desirable scaling with k and the sampling budget N. Combining the advantages of majority voting and BoN, we propose a new inference strategy called Best-of-Majority (BoM), with a pivotal step that restricts the candidates to the responses with high frequency in the N samples before selecting the top-k rewards. We prove that when the sampling budget is N=tildeOmega(C^*), the regret of BoM is O(epsilon_{opt}+epsilon_{mathrm{RM}^2C^*/k}), where C^* is the coverage coefficient, epsilon_{RM} is the estimation error of the reward model, and epsilon_{opt} is the estimation error of reward at the optimal response. We further establish a matching lower bound, certifying that our algorithm is minimax optimal. Beyond optimality, BoM has a key advantage: unlike majority voting and BoN, its performance does not degrade when increasing N. Experimental results of inference on math problems show BoM outperforming both majority voting and BoN.

Question Answering (QA) is a task in which a machine understands a given document and a question to find an answer. Despite impressive progress in the NLP area, QA is still a challenging problem, especially for non-English languages due to the lack of annotated datasets. In this paper, we present the Japanese Question Answering Dataset, JaQuAD, which is annotated by humans. JaQuAD consists of 39,696 extractive question-answer pairs on Japanese Wikipedia articles. We finetuned a baseline model which achieves 78.92% for F1 score and 63.38% for EM on test set. The dataset and our experiments are available at https://github.com/SkelterLabsInc/JaQuAD.

Equivariance theory predicts that an architectural symmetry prior reduces sample complexity by a factor of |G|; this is widely cited but rarely measured as a scaling law with controls that separate the prior from its confounds. On a controlled C_n-symmetric task, we report three findings. First, a wrong-group control with identical orbit size and matched compute is worse than no constraint (joint pairwise CI [+0.79, +3.26] excludes zero, robust across estimators); misaligned constraint is actively harmful, not merely unhelpful. Second, an augmentation baseline equipped with test-time orbit averaging matches the equivariant model exactly -- bit-identical per-epoch validation curves across matched cells -- so the architecture-vs-augmentation gap is conditional on asymmetric test-time computation, not unconditional. Third, the relative exchange rate beta_diff = 1.28 is consistent in sign and order of magnitude with the theoretical 1.0 (single-level CI [+0.92, +2.05]); the more conservative two-level bootstrap (seeds x group sizes) widens this to [-0.63, +1.72], including zero, and a finer-N replication on a sqrt(2)-spaced grid is inconclusive (point estimate -0.82). The methodological contributions -- the relative-rate estimator that cancels the shared-difficulty confound, the wrong-group control, and a pre-specified failure taxonomy -- transfer to any inductive bias whose strength can be parameterised. Honest scoping: the primary estimator beta_diff was adopted post-hoc after the initial analysis revealed a positive-slope identifiability problem; the design was never externally pre-registered; and the headline number rests on an OLS slope over seven group sizes on a coarse N grid. This is an exploratory study, not a confirmatory measurement; the wrong-group result is the cleanest finding and the one we report with the most confidence. A registered replication on fresh seeds is future work.

Benchmarks that measure the forecasting ability of large language models are almost always retrospective: the event has happened, the answer is somewhere on the Web, and the evaluation must defend itself against memorisation. We report the opposite design. Over the 39 days of the 2026 FIFA World Cup, six frontier LLMs -- all with extended thinking and native server-side web search -- were asked before every kickoff, one match at a time, to fill in a seven-market prediction card for all 104 matches, plus 12 group winners and a pre-tournament outright pool; no answer existed when the question was asked, so the evaluation is leakage-free by construction rather than by filtering, and the frozen archive holds 4,494 scored predictions. What the tournament establishes is a set of behaviours the six systems share. On match outcome they average 63.9%, level with backing the bookmaker's favourite -- which is in fact what they usually do. They agree with one another far more often than they are right, so a majority vote adds nothing. They under-commit to draws and to goals, and crowd their scoreline picks onto a single prototypical result. Accuracy tracks how lopsided a fixture is rather than how much is known about it: it collapses in the closest ties, where the dossiers are richest, while questions about the tournament as a whole are answered well. On this task the current generation of frontier systems is not sharply differentiated: the standings hold up at the top and the bottom across the run and churn in the middle, and the margins stay narrow throughout. The briefing dossiers, fixtures and official results are released as a benchmark, together with the scoring code.

We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries. WildBench consists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs. For automated evaluation with WildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo. WildBench evaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments. WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie. Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation. Additionally, we propose a simple method to mitigate length bias, by converting outcomes of ``slightly better/worse'' to ``tie'' if the winner response exceeds the loser one by more than K characters. WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric. WildBench results demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard's 0.91 and AlpacaEval2.0's 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates.