Daily Papers

Voting over multiple LLM responses is a common primitive in test-time scaling and ensemble inference. Collecting more responses can expand the candidate pool and increase the chance that a correct answer is discovered. Under a fixed call budget, a discovered answer still needs to accumulate enough support within the remaining calls to become the final plurality winner, creating a discovery-to-decision gap. In this work, we characterize this gap through the realized vote state and remaining call budget. We derive a sharp recoverability threshold and show that, as sampling proceeds, the observed candidate set can only expand while the set of reachable endpoint winners can only contract, inducing a candidate-level conversion window. Under a specified iid response law, the same state yields exact finite-horizon endpoint probabilities. We further show that merging wrong-answer identities preserves single-call correctness and cannot improve plurality accuracy, and that the effect of redistributing wrong-answer probability depends on the realized vote state. Singleton reachability yields a gold-free exact locking certificate. For a known answer universe, its first trigger is the earliest prefix at which all admissible continuations yield the same fixed-budget output. Empirically, most discovered-but-unselected correct answers lose reachability only after discovery. In a controlled Word16 study, input permutation improves raw-plurality accuracy by 21.1 points with essentially unchanged single-call correctness. Exact locking saves 28-30% of calls at a 16-call budget while preserving every fixed-budget output.

Generating a correct answer does not mean that a language model will select it. We separate factual recall into three steps: generating a correct candidate, ranking the available candidates, and selecting the final answer. Pre-generation readouts predict factual recall and which questions sampling will cover across three model families, but say little about whether an available correct answer will ultimately be selected. Explicit verification with P(True) improves within-question ranking over mean log-likelihood in Gemma, Qwen3, and Llama, with AUROC gains of 0.08--0.12. In a prospectively defined Gemma cohort, verification raises plurality accuracy by about 5 points, and still gains about 2 points over chat-template likelihood, a stronger generative baseline. The advantage is strongest for relations with common-answer priors and depends on access to the entity; masking the entity removes the ranking advantage in larger Qwen models. Finally, the measured benefit depends on how correctness is defined: recall-oriented reference matching can credit option lists favored by likelihood and substantially understate the improvement seen under human semantic judgments. Prior work shows that models can carry latent factual knowledge and judge candidate answers; we show that these capabilities do not collapse into a single notion of ``knowing,'' and trace where information is gained, lost, or mismeasured between availability, ranking, and final choice.

Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems. We introduce PLURAL, a large-scale, value-focused preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries. Using a two-stage generation pipeline, we transform survey responses into synthetic preference triplets that preserve normative value signals while producing realistic scenarios. We release an initial version of PLURAL containing ~500,000 preference triplets representing people in 20 diverse countries. We evaluate PLURAL in three ways: (i) dataset-level validation showing that it preserves both cross-country value differences and within-country diversity from the original survey; (ii) automated evaluation showing that training on PLURAL improves alignment with target countries' cultural profiles, reducing mean absolute error by up to 27.7% relative to strong baselines; and (iii) blind human evaluation with 176 evaluators in India, Brazil, and Japan, who judge PLURAL-aligned responses as more representative of their national values. Together, these results show that PLURAL contains learnable signal for value steering, offering a scalable resource for pluralistic alignment. Dataset: https://huggingface.co/datasets/agdhruv/plural-alignment

Recent debates raised concerns that language models may favor certain viewpoints. But what if the solution is not to aim for a 'view from nowhere' but rather to leverage different viewpoints? We introduce Plurals, a system and Python library for pluralistic AI deliberation. Plurals consists of Agents (LLMs, optionally with personas) which deliberate within customizable Structures, with Moderators overseeing deliberation. Plurals is a generator of simulated social ensembles. Plurals integrates with government datasets to create nationally representative personas, includes deliberation templates inspired by deliberative democracy, and allows users to customize both information-sharing structures and deliberation behavior within Structures. Six case studies demonstrate fidelity to theoretical constructs and efficacy. Three randomized experiments show simulated focus groups produced output resonant with an online sample of the relevant audiences (chosen over zero-shot generation in 75% of trials). Plurals is both a paradigm and a concrete system for pluralistic AI. The Plurals library is available at https://github.com/josh-ashkinaze/plurals and will be continually updated.

We introduce a novel framework for measuring Overton pluralism in LLMs--the extent to which diverse viewpoints are represented in model outputs. We (i) formalize Overton pluralism as a set coverage metric (OvertonScore), (ii) conduct a large-scale U.S.-representative human study (N = 1209; 60 questions; 8 LLMs), and (iii) develop an automated benchmark that closely reproduces human judgments. On average, models achieve OvertonScores of 0.35--0.41, with DeepSeek V3 performing best; yet all models remain far below the theoretical maximum of 1.0, revealing substantial headroom for improvement. Because repeated large-scale human studies are costly and slow, scalable evaluation tools are essential for model development. Hence, we propose an automated benchmark that achieves high rank correlation with human judgments (ρ=0.88), providing a practical proxy without replacing human assessment. By turning pluralistic alignment from a normative aim into a measurable benchmark, our work establishes a foundation for systematic progress toward more pluralistic LLMs.

Large Language Models (LLMs) are increasingly being used to summarize and verify claims in critical sources that inform consequential decisions. Such claims are full of plural definites. If three of five patients improved after taking drug A, the claim the patients improved after drug A is not true, since two did not, yet its negation, the patients didn't improve after drug A, is not true either, since three did. In formal semantics, this truth-value gap is known as the Homogeneity Gap: both the plural definite and its negation are neither true nor false when the property holds for only a fraction of the plural referent. A model that cannot represent this gap turns mixed evidence into a clean verdict, an error that can be costly in high-stakes use. We introduce HomogeneityNLI, a diagnostic dataset of 1,000 minimal pair items crossing determiner, polarity, and group composition across ten logical conditions. Evaluating seven open-weight models, we uncover a pervasive Bivalence Bias: models assign plural definites their homogeneous truth conditions. In the zero-shot setting, no model recognizes the gap for both polarities in more than 15% of scenarios, and negated definites are judged False in 84% of cases. Explicit linguistic rules, chain-of-thought, and enumeration prompts produce unstable, model-specific effects that often entrench the bias. Our findings indicate that current open-weight LLMs acquire the truth conditions of plural predication but not its trivalent logic, and cannot yet be trusted to summarize or verify claims whose evidence is mixed.

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

With increased power and prevalence of AI systems, it is ever more critical that AI systems are designed to serve all, i.e., people with diverse values and perspectives. However, aligning models to serve pluralistic human values remains an open research question. In this piece, we propose a roadmap to pluralistic alignment, specifically using language models as a test bed. We identify and formalize three possible ways to define and operationalize pluralism in AI systems: 1) Overton pluralistic models that present a spectrum of reasonable responses; 2) Steerably pluralistic models that can steer to reflect certain perspectives; and 3) Distributionally pluralistic models that are well-calibrated to a given population in distribution. We also propose and formalize three possible classes of pluralistic benchmarks: 1) Multi-objective benchmarks, 2) Trade-off steerable benchmarks, which incentivize models to steer to arbitrary trade-offs, and 3) Jury-pluralistic benchmarks which explicitly model diverse human ratings. We use this framework to argue that current alignment techniques may be fundamentally limited for pluralistic AI; indeed, we highlight empirical evidence, both from our own experiments and from other work, that standard alignment procedures might reduce distributional pluralism in models, motivating the need for further research on pluralistic alignment.

We present our system for the PlurVA-LLM 2026 Shared Task Track-1, which focuses on pluralistic value alignment in the contexts of China, Indonesia, and Sri Lanka. For this resource-constrained track, we fine-tuned Llama 3.1 8B Instruct using 4-bit QLoRA. Our approach combines option-permutation augmentation for Chinese data, annotator vote expansion for Indonesian data, and binary reformulation with SinhalaMMLU augmentation for Sri Lankan data. We further applied conditional threshold calibration to the predictions for the Sri Lankan data. The final system achieved accuracies of 0.785 for Chinese, 0.715 for Indonesian, and 0.916 for Sri Lankan, resulting in an overall macro-average accuracy of 0.805.

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.

Words of estimative probability (WEP) are expressions of a statement's plausibility (probably, maybe, likely, doubt, likely, unlikely, impossible...). Multiple surveys demonstrate the agreement of human evaluators when assigning numerical probability levels to WEP. For example, highly likely corresponds to a median chance of 0.90+-0.08 in Fagen-Ulmschneider (2015)'s survey. In this work, we measure the ability of neural language processing models to capture the consensual probability level associated to each WEP. Firstly, we use the UNLI dataset (Chen et al., 2020) which associates premises and hypotheses with their perceived joint probability p, to construct prompts, e.g. "[PREMISE]. [WEP], [HYPOTHESIS]." and assess whether language models can predict whether the WEP consensual probability level is close to p. Secondly, we construct a dataset of WEP-based probabilistic reasoning, to test whether language models can reason with WEP compositions. When prompted "[EVENTA] is likely. [EVENTB] is impossible.", a causal language model should not express that [EVENTA&B] is likely. We show that both tasks are unsolved by off-the-shelf English language models, but that fine-tuning leads to transferable improvement.

We have tried to reproduce the results of the paper "Natural Language Inference over Interaction Space" submitted to ICLR 2018 conference as part of the ICLR 2018 Reproducibility Challenge. Initially, we were not aware that the code was available, so we started to implement the network from scratch. We have evaluated our version of the model on Stanford NLI dataset and reached 86.38% accuracy on the test set, while the paper claims 88.0% accuracy. The main difference, as we understand it, comes from the optimizers and the way model selection is performed.

The rapid advancement of language models (LMs) necessitates robust alignment with diverse user values. However, current preference optimization approaches often fail to capture the plurality of user opinions, instead reinforcing majority viewpoints and marginalizing minority perspectives. We introduce PERSONA, a reproducible test bed designed to evaluate and improve pluralistic alignment of LMs. We procedurally generate diverse user profiles from US census data, resulting in 1,586 synthetic personas with varied demographic and idiosyncratic attributes. We then generate a large-scale evaluation dataset containing 3,868 prompts and 317,200 feedback pairs obtained from our synthetic personas. Leveraging this dataset, we systematically evaluate LM capabilities in role-playing diverse users, verified through human judges, and the establishment of both a benchmark, PERSONA Bench, for pluralistic alignment approaches as well as an extensive dataset to create new and future benchmarks. The full dataset and benchmarks are available here: https://www.synthlabs.ai/research/persona.

Users increasingly rely on large language models (LLMs) to analyse collections of independent inputs, such as product reviews or social media posts, and return a single aggregated answer. This setting requires multi-instance processing (MIP), where a model combines information from multiple instances into one prediction. We study a task in which the model must infer a binary label for each instance, and identify the most frequent label among the full instance set. The correct answer should be invariant to instance order. By evaluating fifteen LLMs across five datasets and seven instance order patterns, with each instance set held fixed across patterns, we find substantial order sensitivity. Accuracy can differ by about 90 percentage points between the best and worst order patterns, with models generally more accurate when majority-label instances appear at the start than when they appear in the middle or at the end. For subsets of models, available reasoning traces and gradient-based attribution reveal uneven emphasis across positions. Traces disproportionately mention early instances, while attribution scores are generally higher near the beginning and end than in the middle. These results show that instance order can systematically bias aggregated predictions, highlighting the need to evaluate MIP across different input arrangements.

Plan accuracy alone cannot show whether a speech-delivery decision follows its source: a fixed prior may match the original label yet fail to respond appropriately when a cue changes. VoxReason provides a 100-case verifier benchmark that holds each utterance fixed, edits one designated source-label cue, and scores cited evidence, eight plan fields, and the permitted response. On a source-key-disjoint test of 24 cases, a source-emotion prior reaches plan-slot accuracy 0.958, but none of the 24 edited neutral targets appears in its training labels; its required-change accuracy is 0.000. This diagnoses the support boundary of this prior, not its performance on supported edits. In a complementary 32-case emotion-disjoint test, the prior has seen all edited neutral targets but neither original test emotion; its plan-slot accuracy is 0.219 and required-change accuracy is 1.000. The partitions reuse and overlap the same 100 cases, so these deterministic diagnostics are not independent cohorts or learned-planner results. The benchmark evaluates derived labels and structured plans, not audio input, generated speech, or listener judgments.

The BITEM team entered both subtasks of the NTCIR-19 R2C2 task with a single agentic pipeline, in which a model searches, reads and records evidence over a movie corpus while an orchestrator holds the record and rules on what may be submitted. A claim is admitted only once an entailment cascade has checked it against the passage it cites, and an answer is released only once enough checked evidence stands behind it. Each question is run three or four times, every pass retrieving from a corpus stripped of what the earlier passes have already seen. The confidence filed with each answer is computed by the orchestrator from what the run leaves behind and is never asked of the model, which is offered no way to rate itself. The two retrieval runs placed 4th and 5th of 22, pooling the passes was worth 0.0709 nDCG@20, and the gain was largest on the multi-hop and post-processing-heavy questions, where the organisers rank the pooled run top of the field. Sixteen of the 25 answer runs were built on passages these two runs supplied, 12 of them filed by other teams. HMR rewards a system whose confidence is high where it answers right and low where it answers wrong. The pipeline reached an accuracy of 0.9219, 6th of 25, while the confidence filed with those answers gave an HMR of 0.4915, 13th. A few rules crafted over those same recorded signals, with no further model call and no further retrieval, raise that to an accuracy of 0.9375, 5th, and an HMR of 0.6985, 9th. Ranking on HMR alone can reward a system for answering wrongly with low confidence, so we propose accHMR, the accuracy multiplied by HMR, which reports the reward in proportion to the accuracy, and on which the revised rules would have scored 0.6549, 5th. For future work, fitting a model on the numbers the pipeline already produces, rather than writing such rules by hand, would be a real step forward.

GPT-3 can perform numerous tasks when provided a natural language prompt that contains a few training examples. We show that this type of few-shot learning can be unstable: the choice of prompt format, training examples, and even the order of the training examples can cause accuracy to vary from near chance to near state-of-the-art. We demonstrate that this instability arises from the bias of language models towards predicting certain answers, e.g., those that are placed near the end of the prompt or are common in the pre-training data. To mitigate this, we first estimate the model's bias towards each answer by asking for its prediction when given the training prompt and a content-free test input such as "N/A". We then fit calibration parameters that cause the prediction for this input to be uniform across answers. On a diverse set of tasks, this contextual calibration procedure substantially improves GPT-3 and GPT-2's average accuracy (up to 30.0% absolute) and reduces variance across different choices of the prompt.

Many people are interested in ChatGPT since it has become a prominent AIGC model that provides high-quality responses in various contexts, such as software development and maintenance. Misuse of ChatGPT might cause significant issues, particularly in public safety and education, despite its immense potential. The majority of researchers choose to publish their work on Arxiv. The effectiveness and originality of future work depend on the ability to detect AI components in such contributions. To address this need, this study will analyze a method that can see purposely manufactured content that academic organizations use to post on Arxiv. For this study, a dataset was created using physics, mathematics, and computer science articles. Using the newly built dataset, the following step is to put originality.ai through its paces. The statistical analysis shows that Originality.ai is very accurate, with a rate of 98%.

A decision model can give probabilities that sum to one for every question yet disagree with itself when the same decision is broken into smaller steps. We study this form of probabilistic coherence in Jev and the English Laya checkpoint, using 2,500 matched examples per system across TREC, CLINC150, and MASSIVE. Across 72,000 classification questions, we compare direct fine-label predictions with broad-category probabilities and predictions reconstructed through those categories. Both systems show substantial disagreement: mean category-level total variation ranges from 0.219 to 0.349 for Jev and from 0.424 to 0.689 for Laya, on a scale where zero means exact agreement. The consequences differ sharply. On CLINC150, reconstruction reduces Jev's accuracy by 22.9 percentage points (paired 95% bootstrap interval: [-24.9, -20.9]) and improves Laya's by 21.3 points ([18.0, 24.5]). The same directions hold across all three datasets, with all six unadjusted accuracy-change intervals excluding zero. Improved accuracy can also accompany less reliable confidence: on MASSIVE, Laya gains 9.2 accuracy points while its expected calibration error rises from 0.046 to 0.124. Error analysis identifies both broad-category mistakes and within-category confusions. These findings show why decision systems need joint evaluation of accuracy, confidence calibration, and probability coherence in the workflow used by an application.

Large language models can arrive at the same answer through reasoning paths that are unstable, contradictory, or difficult to rank consistently -- a failure mode especially prevalent in multi-step deductive reasoning. Existing methods assess reliability primarily through output dispersion -- measuring how much sampled answers differ -- but this discards a complementary signal: whether the model can consistently rank competing reasoning candidates. We propose structural uncertainty, a consistency-aware framework derived from the stability of self-preference-induced rankings over sampled reasoning solutions. Given a query, we generate multiple candidate solutions and ask the model to judge pairwise preferences among its own outputs. We aggregate self-preferences into ranking distributions via Bradley-Terry modeling with PageRank, and decompose the signal into two entropy-based components: across-trial ranking instability and within-trial candidate ambiguity. Across five LLMs and eight benchmarks, structural signals provide information complementary to answer dispersion: on logical and mathematical reasoning tasks, the combination improves identification of unreliable instances, while on factual retrieval the structural signal collapses toward uniformity, diagnosing a regime boundary where reasoning-level consistency evaluation is uninformative. The two components relate differently to accuracy: within-trial ambiguity correlates positively with correctness -- consistent with settings where multiple plausible solution paths remain competitive -- while across-trial instability correlates negatively, signaling unreliable reasoning. Structural uncertainty is best understood not as a universal confidence estimator, but as a regime-sensitive evaluator of logical reasoning consistency.

Sampling based consistency is widely used for hallucination detection, yet aggregate performance can conceal systematic differences in which errors are detectable. This work studies that heterogeneity across four language models and three factual question answering datasets. Partitioning hallucinations by answer agreement reveals high agreement (Ghost) and low agreement (Flickering) regimes with an apparent detectability gap of 0.35 to 0.46 AUC. Because the statistics used to define the regimes and measure this gap are strongly coupled (|ρ|approx0.94 to 1.00), the raw result is treated as a property of agreement based detection rather than independent evidence. After freezing regime assignments, lexical and semantic response dispersion preserve the asymmetry, with bootstrap 95% intervals excluding zero in all 12 model and dataset settings. A stricter test using individual diffusion trajectories and no cross seed information preserves the asymmetry across all three LLaDA datasets (p<0.005) and directionally across all three Dream datasets, with one reaching significance. The hard regime varies substantially in prevalence across models (16% to 77%), and matched prompts frequently change regimes between models. These findings show that aggregate detection metrics conceal persistent, model dependent heterogeneity in language model failures and motivate regime conditioned evaluation.

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

We propose a new test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more. To attain high accuracy on this test, models must possess extensive world knowledge and problem solving ability. We find that while most recent models have near random-chance accuracy, the very largest GPT-3 model improves over random chance by almost 20 percentage points on average. However, on every one of the 57 tasks, the best models still need substantial improvements before they can reach expert-level accuracy. Models also have lopsided performance and frequently do not know when they are wrong. Worse, they still have near-random accuracy on some socially important subjects such as morality and law. By comprehensively evaluating the breadth and depth of a model's academic and professional understanding, our test can be used to analyze models across many tasks and to identify important shortcomings.

In Grammatical Error Correction, systems are evaluated by the number of errors they correct. However, no one has assessed whether all error types are equally important. We provide and apply a method to quantify the importance of different grammatical error types to humans. We show that some rare errors are considered disturbing while other common ones are not. This affects possible directions to improve both systems and their evaluation.

Incongruence detection in eyewitness narratives is critical for understanding the reliability of testimonies, yet traditional approaches often fail to address the nuanced inconsistencies inherent in such accounts. In this paper, we introduce a novel task of incongruence detection in eyewitness testimonies. Given a pair of testimonies containing of multiple pairs of question and answer by two subjects, we identify contextually related incongruence between the two subjects. We also mark the span of incongruences in the utterances. To achieve this, we developed MIND(MultI-EyewitNess Deception) - a comprehensive dataset consisting of 2927 pairs of contextually related answers designed to capture both explicit and implicit contradictions. INstruction - TunEd iNcongruity Detection framework based on 6W and multi-hop reasoning approach, aka. INTEND. Drawing from investigative techniques, INTEND address the task as a close-style problem, contradicting on the who, what, when, where and why aspect of the content. Our findings shows that prompt tuning, especially when utilizing our framework, enhances the detection of incongruences by a margin of +5.63 percent. We compare our approach with multiple fine-tuning and prompt tuning techniques on MLMs and LLMs. Emperical results demonstrate convincing performance improvement in F1-score over fine-tuned and regular prompt-tuning techniques, highlighting the effectiveness of our approach.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

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.

We present an alternative method of evaluating Quality Estimation systems, which is based on a linguistically-motivated Test Suite. We create a test-set consisting of 14 linguistic error categories and we gather for each of them a set of samples with both correct and erroneous translations. Then, we measure the performance of 5 Quality Estimation systems by checking their ability to distinguish between the correct and the erroneous translations. The detailed results are much more informative about the ability of each system. The fact that different Quality Estimation systems perform differently at various phenomena confirms the usefulness of the Test Suite.

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

Agentic retrieval-augmented reasoning pipelines are increasingly used to structure how large language models (LLMs) incorporate external evidence in clinical decision support. These systems iteratively retrieve curated domain knowledge and synthesize it into structured reports before answer selection. Although such pipelines can improve performance, their impact on reliability under model variability remains unclear. In real-world deployment, heterogeneous models may align, diverge, or synchronize errors in ways not captured by accuracy. We evaluated 34 LLMs on 169 expert-curated publicly available radiology questions, comparing zero-shot inference with a radiology-specific multi-step agentic retrieval condition in which all models received identical structured evidence reports derived from curated radiology knowledge. Agentic inference reduced inter-model decision dispersion (median entropy 0.48 vs. 0.13) and increased robustness of correctness across models (mean 0.74 vs. 0.81). Majority consensus also increased overall (P<0.001). Consensus strength and robust correctness remained correlated under both strategies (ho=0.88 for zero-shot; ho=0.87 for agentic), although high agreement did not guarantee correctness. Response verbosity showed no meaningful association with correctness. Among 572 incorrect outputs, 72% were associated with moderate or high clinically assessed severity, although inter-rater agreement was low (appa=0.02). Agentic retrieval therefore was associated with more concentrated decision distributions, stronger consensus, and higher cross-model robustness of correctness. These findings suggest that evaluating agentic systems through accuracy or agreement alone may not always be sufficient, and that complementary analyses of stability, cross-model robustness, and potential clinical impact are needed to characterize reliability under model variability.

In this paper, we have shown a method of improving the quality of neural machine translation by translating/transliterating name entities as a preprocessing step. Through experiments we have shown the performance gain of our system. For evaluation we considered three types of name entities viz person names, location names and organization names. The system was able to correctly translate mostly all the name entities. For person names the accuracy was 99.86%, for location names the accuracy was 99.63% and for organization names the accuracy was 99.05%. Overall, the accuracy of the system was 99.52%

A new generation of AI models generates step-by-step reasoning text before producing an answer. This text appears to offer a human-readable window into their computation process, and is increasingly relied upon for transparency and interpretability. However, it is unclear whether human understanding of this text matches the model's actual computational process. In this paper, we investigate a necessary condition for correspondence: the ability of humans to identify which steps in a reasoning text causally influence later steps. We evaluated humans on this ability by composing questions based on counterfactual measurements and found a significant discrepancy: participant accuracy was only 29.3%, barely above chance (25%), and remained low (42%) even when evaluating the majority vote on questions with high agreement. Our results reveal a fundamental gap between how humans interpret reasoning texts and how models use it, challenging its utility as a simple interpretability tool. We argue that reasoning texts should be treated as an artifact to be investigated, not taken at face value, and that understanding the non-human ways these models use language is a critical research direction.

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

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

Benchmarks measure whether a model is correct. They do not measure whether a model is reliable. This distinction is largely academic for single-shot inference, but becomes critical for agentic AI systems, where a single rephrased prompt can trigger cascading failures in multi-step execution. Yet this form of instability is not captured by existing evaluations. Hallucinations live in variance: they arise when semantically equivalent prompts activate inconsistent internal pathways, producing divergent outputs. Consistent but incorrect outputs reflect bias or missing knowledge; confident guessing reflects calibration failure. Neither constitutes hallucination under this definition. When error is variance-dominated, reducing redundant pathways improves reliability without adding knowledge. We formalize this through Semantic Stability (SS), measured via Paraphrase Consistency (PC@k): generate k paraphrases, greedy decode each, compute mode agreement. SS is a diagnostic for variance-driven unreliability, not a method for improving correctness. We show that a dense Qwen3-0.6B agrees with itself only 23.8% of the time; at 32% sparsity, agreement jumps to 55.9%. A phase diagram reveals the sweet spot where variance reduction outpaces bias accumulation, and regimes where stability collapses onto wrong answers.

Context: Code reviews are crucial for software quality. Recent AI advances have allowed large language models (LLMs) to review and fix code; now, there are tools that perform these reviews. However, their reliability and accuracy have not yet been systematically evaluated. Objective: This study compares different LLMs' performance in detecting code correctness and suggesting improvements. Method: We tested GPT4o and Gemini 2.0 Flash on 492 AI generated code blocks of varying correctness, along with 164 canonical code blocks from the HumanEval benchmark. To simulate the code review task objectively, we expected LLMs to assess code correctness and improve the code if needed. We ran experiments with different configurations and reported on the results. Results: With problem descriptions, GPT4o and Gemini 2.0 Flash correctly classified code correctness 68.50% and 63.89% of the time, respectively, and corrected the code 67.83% and 54.26% of the time for the 492 code blocks of varying correctness. Without problem descriptions, performance declined. The results for the 164 canonical code blocks differed, suggesting that performance depends on the type of code. Conclusion: LLM code reviews can help suggest improvements and assess correctness, but there is a risk of faulty outputs. We propose a process that involves humans, called the "Human in the loop LLM Code Review" to promote knowledge sharing while mitigating the risk of faulty outputs.

The paper analyzes the accuracy of publicly available object-recognition systems on a geographically diverse dataset. This dataset contains household items and was designed to have a more representative geographical coverage than commonly used image datasets in object recognition. We find that the systems perform relatively poorly on household items that commonly occur in countries with a low household income. Qualitative analyses suggest the drop in performance is primarily due to appearance differences within an object class (e.g., dish soap) and due to items appearing in a different context (e.g., toothbrushes appearing outside of bathrooms). The results of our study suggest that further work is needed to make object-recognition systems work equally well for people across different countries and income levels.

A retrieval-augmented QA system can return different answers after an index expansion even when its requested model identifier, prompt, retrieval policy, evidence depth, rendering, and exposed generation controls are held fixed. Aggregate accuracy may hide these changes when gains and losses cancel, while ordinary generation variability makes one-shot comparisons overstate update effects. We call the hidden phenomenon accuracy-blind answer churn and introduce the Snapshot Compatibility Audit, which estimates excess answer churn by subtracting same-snapshot repeat disagreement from cross-snapshot disagreement. We instantiate it by expanding one frozen FineWeb prefix from one to seven shards. In a preregistered 400-question Natural Questions study, normalized-exact and blinded-semantic excess churn are 6.44 and 10.25 percentage points while exact-match accuracy changes by only -1.50 points. A post-hoc analysis finds repeat-stable semantic flips on 40/400 questions. A separately preregistered 200-question TriviaQA study yields smaller, directionally consistent excess churn while exact-match accuracy moves in the opposite direction. An outcome-blind post-hoc 100-question subset replication with a second DeepSeek generator and serving configuration finds 8.75 pp of semantic excess churn even as exact match rises by 3.00 percentage points. Answer-level compatibility can therefore fail without a conspicuous or consistently directed utility shift. Retrieval-augmented releases should audit compatibility alongside utility.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

There is a growing literature on reasoning by large language models (LLMs), but the discussion on the uncertainty in their responses is still lacking. Our aim is to assess the extent of confidence that LLMs have in their answers and how it correlates with accuracy. Confidence is measured (i) qualitatively in terms of persistence in keeping their answer when prompted to reconsider, and (ii) quantitatively in terms of self-reported confidence score. We investigate the performance of three LLMs -- GPT4o, GPT4-turbo and Mistral -- on two benchmark sets of questions on causal judgement and formal fallacies and a set of probability and statistical puzzles and paradoxes. Although the LLMs show significantly better performance than random guessing, there is a wide variability in their tendency to change their initial answers. There is a positive correlation between qualitative confidence and accuracy, but the overall accuracy for the second answer is often worse than for the first answer. There is a strong tendency to overstate the self-reported confidence score. Confidence is only partially explained by the underlying token-level probability. The material effects of prompting on qualitative confidence and the strong tendency for overconfidence indicate that current LLMs do not have any internally coherent sense of confidence.

To help accelerate progress in multi-target, multi-camera tracking systems, we present (i) a new pair of precision-recall measures of performance that treats errors of all types uniformly and emphasizes correct identification over sources of error; (ii) the largest fully-annotated and calibrated data set to date with more than 2 million frames of 1080p, 60fps video taken by 8 cameras observing more than 2,700 identities over 85 minutes; and (iii) a reference software system as a comparison baseline. We show that (i) our measures properly account for bottom-line identity match performance in the multi-camera setting; (ii) our data set poses realistic challenges to current trackers; and (iii) the performance of our system is comparable to the state of the art.

Retrieval augmented generation (RAG) is often used to fix hallucinations and provide up-to-date knowledge for large language models (LLMs). However, in cases when the LLM alone incorrectly answers a question, does providing the correct retrieved content always fix the error? Conversely, in cases where the retrieved content is incorrect, does the LLM know to ignore the wrong information, or does it recapitulate the error? To answer these questions, we systematically analyze the tug-of-war between a LLM's internal knowledge (i.e. its prior) and the retrieved information in settings when they disagree. We test GPT-4 and other LLMs on question-answering abilities across datasets with and without reference documents. As expected, providing the correct retrieved information fixes most model mistakes (94% accuracy). However, when the reference document is perturbed with increasing levels of wrong values, the LLM is more likely to recite the incorrect, modified information when its internal prior is weaker but is more resistant when its prior is stronger. Similarly, we also find that the more the modified information deviates from the model's prior, the less likely the model is to prefer it. These results highlight an underlying tension between a model's prior knowledge and the information presented in reference documents.

Can humans tell whether a news article was written by a person or a large language model (LLM)? We investigate this question using JudgeGPT, a study platform that independently measures source attribution (human vs. machine) and authenticity judgment (legitimate vs. fake) on continuous scales. From 2,318 judgments collected from 1,054 participants across content generated by six LLMs, we report five findings: (1) participants cannot reliably distinguish machine-generated from human-written text (p > .05, Welch's t-test); (2) this inability holds across all tested models, including open-weight models with as few as 7B parameters; (3) self-reported domain expertise predicts judgment accuracy (r = .35, p < .001) whereas political orientation does not (r = -.10, n.s.); (4) clustering reveals distinct response strategies ("Skeptics" vs. "Believers"); and (5) accuracy degrades after approximately 30 sequential evaluations due to cognitive fatigue. The answer, in short, is no: humans cannot reliably tell. These results indicate that user-side detection is not a viable defense and motivate system-level countermeasures such as cryptographic content provenance.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

Precision and Recall are two prominent metrics of generative performance, which were proposed to separately measure the fidelity and diversity of generative models. Given their central role in comparing and improving generative models, understanding their limitations are crucially important. To that end, in this work, we identify a critical flaw in the common approximation of these metrics using k-nearest-neighbors, namely, that the very interpretations of fidelity and diversity that are assigned to Precision and Recall can fail in high dimensions, resulting in very misleading conclusions. Specifically, we empirically and theoretically show that as the number of dimensions grows, two model distributions with supports at equal point-wise distance from the support of the real distribution, can have vastly different Precision and Recall regardless of their respective distributions, hence an emergent asymmetry in high dimensions. Based on our theoretical insights, we then provide simple yet effective modifications to these metrics to construct symmetric metrics regardless of the number of dimensions. Finally, we provide experiments on real-world datasets to illustrate that the identified flaw is not merely a pathological case, and that our proposed metrics are effective in alleviating its impact.

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.

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

Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with an LLM matcher. However, eliminating positional influence entirely still does not reliably yield accuracy gains, while cyclic permutation often improves them. For two-stage prompting, an aggregate measure of recall imbalance and a direct per-question measure of order sensitivity both fail to show reliable debiasing.

Multi-agent debate, where teams of LLMs iteratively exchange rationales and vote on answers, is widely deployed under the assumption that peer review filters hallucinations. Yet the failure dynamics of homogeneous debate remain poorly understood, therefore we report findings from a controlled empirical study of teams of N{=}10 homogeneous agents (Qwen2.5-7B, Llama-3.1-8B, Ministral-3-8B) across R{=}3 debate rounds on two high-difficulty benchmarks (GSM-Hard and MMLU-Hard). We compare peer debate against isolated self-correction and a stochastic noise control that injects rationales from unrelated problems. We decompose debate failure into three model-dependent pathways: sycophantic conformity, where agents uncritically adopt majority answers (modal adoption up to 85.5%); contextual fragility, where peer rationales destabilize previously correct reasoning (vulnerability rate up to 70.0%); and consensus collapse, where plurality voting discards correct answers already present in the generation pool (oracle gap up to 32.3 percentage points). Ablations over communication density (K in {2,4,9}) and sampling temperature (T in {0.4, 0.7}) show that conformity reaches high levels at minimal peer exposure (K{=}2) and intensifies with greater initial diversity. Across all configurations, debate consumes 2.1-3.4times more tokens (up to 28,631 tokens per problem) than self-correction for equal or lower accuracy. Our results indicate that, within the 7-8B parameter class, homogeneous teams without structured roles do not benefit from unguided peer exchange, and that isolated self-correction consistently offers a more favorable cost-accuracy tradeoff.

Language models often receive a question together with a claim about what another source answered. We audit whether such claims destabilize answers in multiple-choice question answering. For each item, we hold one wrong option fixed across misleading conditions and vary the cue template attached to it. We introduce neutral-conditioned misleading cue adoption rate (NC-MCAR), which measures switches to that option only on valid cued trials where the same model first selected the gold answer under a neutral prompt. This is a measure of answer instability, not proof that the model knew the answer or that all deference is irrational. We evaluate four instruction-following models on MMLU-Pro and IndicMMLU-Pro in English, Hindi, Bengali, Tamil, and Telugu. Across 220{,}000 outputs, the expert template yields 41.1\% aggregate NC-MCAR, compared with 12.5\% for the majority template. These two conditions use the same wrong option and final instruction. Filler accuracy remains well above expert-wrong accuracy, while correct-cue prompts have high valid-response accuracy. The audit documents answer instability relevant to grounding under the tested forced-choice prompts: a bare, unverified source claim can outweigh an answer that was previously consistent with the task evidence.

Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.

In this paper, we uncover a systematic bias in the evaluation paradigm of adopting large language models~(LLMs), e.g., GPT-4, as a referee to score and compare the quality of responses generated by candidate models. We find that the quality ranking of candidate responses can be easily hacked by simply altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other, e.g., Vicuna-13B could beat ChatGPT on 66 over 80 tested queries with ChatGPT as an evaluator. To address this issue, we propose a calibration framework with three simple yet effective strategies: 1) Multiple Evidence Calibration, which requires the evaluator model to generate multiple evaluation evidence before assigning ratings; 2) Balanced Position Calibration, which aggregates results across various orders to determine the final score; 3) Human-in-the-Loop Calibration, which introduces a balanced position diversity entropy to measure the difficulty of each example and seeks human assistance when needed. We also manually annotate the "win/tie/lose" outcomes of responses from ChatGPT and Vicuna-13B in the Vicuna Benchmark's question prompt, and extensive experiments demonstrate that our approach successfully mitigates evaluation bias, resulting in closer alignment with human judgments. We release our code and human annotation at https://github.com/i-Eval/FairEval to facilitate future research.

Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity. To understand this inconsistency, we unify existing methods into a standard framework, showing they are equivalent to solving a common optimization problem: minimize average loss subject to a method-specific mixing law -- an implicit assumption on the relationship between loss and mixture proportions. This framework suggests that measuring the fidelity of a method's mixing law can offer insights into its performance. Empirically, we find that existing methods set their mixing law parameters inaccurately, resulting in the inconsistent mixing performance we observe. Using this insight, we derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.27 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.012 test perplexity points.

Self-play is an effective paradigm for language-model self-evolution, but without guidance, solver performance can plateau or decline across rounds. Unguided methods steer question generation with signals such as difficulty, learnability, or diversity. These signals keep questions challenging and varied but do not specify which unresolved reasoning weaknesses later rounds should target. Guided methods obtain direction from external task resources, including human examples, document corpora, or specified difficulty targets, and therefore rely on task information supplied outside the self-play loop. We show that the needed direction can instead be derived from the solver's own failure history. We introduce DiagEvo, whose diagnostician extracts recurring error causes from this history and stores them in a hierarchical error-cause memory. The memory groups related causes under skill nodes and tracks each as Active or Mastered according to self-consistency on targeted questions. The challenger uses these states and recurrence counts to balance cause-targeted generation with free exploration. Double-confidence filtering retains intermediate-difficulty questions only when the most common solver answer has a clear vote lead. DiagEvo derives its curriculum from information produced during self-play, without external task resources. With the default 4B diagnostician, DiagEvo outperforms every baseline in mean accuracy across all nine benchmarks for each of the three solvers: Qwen3-4B, Qwen3-8B, and OctoThinker-8B. On Qwen3-8B, it reaches 72.3% mean accuracy across five mathematical reasoning benchmarks, 4.5 percentage points above R-Zero. Its mean accuracy across all nine benchmarks is 57.4%, 1.1 percentage points above DARC. Ablations show that the hierarchical error-cause memory and double-confidence filtering both contribute to these gains.

Robust, faithful and harm-free pronoun use for individuals is an important goal for language models as their use increases, but prior work tends to study only one or two of these characteristics at a time. To measure progress towards the combined goal, we introduce the task of pronoun fidelity: given a context introducing a co-referring entity and pronoun, the task is to reuse the correct pronoun later. We present RUFF, a carefully-designed dataset of over 5 million instances to measure robust pronoun fidelity in English, and we evaluate 37 popular large language models across architectures (encoder-only, decoder-only and encoder-decoder) and scales (11M-70B parameters). When an individual is introduced with a pronoun, models can mostly faithfully reuse this pronoun in the next sentence, but they are significantly worse with she/her/her, singular they and neopronouns. Moreover, models are easily distracted by non-adversarial sentences discussing other people; even one additional sentence with a distractor pronoun causes accuracy to drop on average by 34%. Our results show that pronoun fidelity is neither robust, nor due to reasoning, in a simple, naturalistic setting where humans achieve nearly 100% accuracy. We encourage researchers to bridge the gaps we find and to carefully evaluate reasoning in settings where superficial repetition might inflate perceptions of model performance.

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

Pretrained Language Models (LMs) have demonstrated ability to perform numerical reasoning by extrapolating from a few examples in few-shot settings. However, the extent to which this extrapolation relies on robust reasoning is unclear. In this paper, we investigate how well these models reason with terms that are less frequent in the pretraining data. In particular, we examine the correlations between the model performance on test instances and the frequency of terms from those instances in the pretraining data. We measure the strength of this correlation for a number of GPT-based language models (pretrained on the Pile dataset) on various numerical deduction tasks (e.g., arithmetic and unit conversion). Our results consistently demonstrate that models are more accurate on instances whose terms are more prevalent, in some cases above 70% (absolute) more accurate on the top 10\% frequent terms in comparison to the bottom 10\%. Overall, although LMs exhibit strong performance at few-shot numerical reasoning tasks, our results raise the question of how much models actually generalize beyond pretraining data, and we encourage researchers to take the pretraining data into account when interpreting evaluation results.

Many modern AI question-answering systems convert text into vectors and retrieve the closest matches to a user question. While effective for topical similarity, similarity scores alone do not explain why some retrieved text can serve as evidence while other equally similar text cannot. When many candidates receive similar scores, systems may select sentences that are redundant, incomplete, or address different conditions than the question requires. This paper presents a deterministic evidence selection framework for retrieval-augmented question answering. The approach introduces Meaning-Utility Estimation (MUE) and Diversity-Utility Estimation (DUE), fixed scoring and redundancy-control procedures that determine evidence admissibility prior to answer generation. Each sentence or record is evaluated independently using explicit signals for semantic relatedness, term coverage, conceptual distinctiveness, and redundancy. No training or fine-tuning is required. In the prototype, a unit is accepted only if it explicitly states the fact, rule, or condition required by the task. Units are not merged or expanded. If no unit independently satisfies the requirement, the system returns no answer. This deterministic gating produces compact, auditable evidence sets and establishes a clear boundary between relevant text and usable evidence.

Language models gain accuracy from more compute at test time, but majority voting over independent samples saturates: as samples grow, the vote converges to the model's most frequent answer. Communication can add what sampling cannot: an agent that solves a problem can pass the key step to the others. We present TalkMesh, a decentralized mesh of small language model agents that learns when and what to communicate. Each agent samples a proposal and scores it with a trained confidence head. The most confident agent broadcasts a hint; agents below a confidence threshold revise, keeping each revision that outscores its proposal. Gossip consensus approximates the vote weighted by confidence without a coordinator. A talk policy, trained with group relative policy optimization on the change in correctness after revision, writes hints and revisions. With three agents, which together generate at most six outputs, the mesh reaches the accuracy of majority voting over 32 samples with each of three models. Trained with at most 8 agents and evaluated with 32, it raises accuracy from 0.568 under self-consistency to 0.705 (Qwen3.5-0.8B, GSM8K) and from 0.492 to 0.722 (SmolLM3-3B, MATH-500). When 4 of 8 agents collude on a wrong answer with fabricated confidence and poisoned hints, majority vote accuracy falls to 0.000 (Qwen3.5-0.8B, GSM8K). A defended mesh, whose agents rescore solutions with their own confidence heads, retains 0.507. Across reasoning, embodied coordination, and traffic signal control, messages improve a decision when the acting agent cannot observe the information it requires and another agent can send it.

As LLM-based agents are deployed in production systems, understanding their behavioral consistency (whether they produce similar action sequences when given identical tasks) becomes critical for reliability. We study consistency in the context of SWE-bench, a challenging software engineering benchmark requiring complex, multi-step reasoning. Comparing Claude~4.5~Sonnet, GPT-5, and Llama-3.1-70B across 50 runs each (10 tasks times 5 runs), we find that across models, higher consistency aligns with higher accuracy: Claude achieves the lowest variance (CV: 15.2\%) and highest accuracy (58\%), GPT-5 is intermediate (CV: 32.2\%, accuracy: 32\%), and Llama shows the highest variance (CV: 47.0\%) with lowest accuracy (4\%). However, within a model, consistency can amplify both correct and incorrect interpretations. Our analysis reveals a critical nuance: consistency amplifies outcomes rather than guaranteeing correctness. 71\% of Claude's failures stem from "consistent wrong interpretation": making the same incorrect assumption across all runs. Interestingly, GPT-5 achieves similar early strategic agreement as Claude (diverging at step 3.4 vs.\ 3.2) but exhibits 2.1times higher variance, suggesting that divergence timing alone does not determine consistency. These findings suggest that for production deployment, interpretation accuracy matters more than execution consistency, with implications for agent evaluation and training.

Large language models are increasingly used for content moderation, but most evaluations still report aggregate accuracy on individual benchmarks. We introduce Safety-Flag, which places seven widely used safety benchmarks (BeaverTails, XSTest, Ethics, WildGuard, Aegis, ToxiChat, and ToxiGen) into a single balanced flag / do-not-flag protocol. We release item-level decisions and confidence scores for six general-purpose LLMs and four dedicated guards, together with three reference models, evaluated on the same items. Safety-Flag measures three dimensions of moderator reliability: error direction, probability calibration, and confidence-based error ranking for human review. They often disagree. Aggregate accuracy does not reveal error direction: one model flags 85% of benign content, whereas another misses 54% of harmful content. All six general-purpose models are overconfident; fitting one temperature per model reduces calibration error by 2.8--6.0times without changing predicted labels or confidence ordering. Confidence-based abstention lowers selective risk for every model, although the gains depend on how well confidence ranks errors. Dedicated guards produce fewer false alarms and are better calibrated, but several have higher miss rates outside their documented coverage. We release the benchmark, fixed item lists, evaluation code, per-item model outputs, and leaderboard at: https://github.com/yibo-hu-lab/safety-flag-benchmark.

Recent work has highlighted the culturally-contingent nature of commonsense knowledge. We introduce AMAMMER{epsilon}, a test set of 525 multiple-choice questions designed to evaluate the commonsense knowledge of English LLMs, relative to the cultural contexts of Ghana and the United States. To create AMAMMER{epsilon}, we select a set of multiple-choice questions (MCQs) from existing commonsense datasets and rewrite them in a multi-stage process involving surveys of Ghanaian and U.S. participants. In three rounds of surveys, participants from both pools are solicited to (1) write correct and incorrect answer choices, (2) rate individual answer choices on a 5-point Likert scale, and (3) select the best answer choice from the newly-constructed MCQ items, in a final validation step. By engaging participants at multiple stages, our procedure ensures that participant perspectives are incorporated both in the creation and validation of test items, resulting in high levels of agreement within each pool. We evaluate several off-the-shelf English LLMs on AMAMMER{epsilon}. Uniformly, models prefer answers choices that align with the preferences of U.S. annotators over Ghanaian annotators. Additionally, when test items specify a cultural context (Ghana or the U.S.), models exhibit some ability to adapt, but performance is consistently better in U.S. contexts than Ghanaian. As large resources are devoted to the advancement of English LLMs, our findings underscore the need for culturally adaptable models and evaluations to meet the needs of diverse English-speaking populations around the world.

Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.

Human values are crucial to human decision-making. Value pluralism is the view that multiple correct values may be held in tension with one another (e.g., when considering lying to a friend to protect their feelings, how does one balance honesty with friendship?). As statistical learners, AI systems fit to averages by default, washing out these potentially irreducible value conflicts. To improve AI systems to better reflect value pluralism, the first-order challenge is to explore the extent to which AI systems can model pluralistic human values, rights, and duties as well as their interaction. We introduce ValuePrism, a large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations. ValuePrism's contextualized values are generated by GPT-4 and deemed high-quality by human annotators 91% of the time. We conduct a large-scale study with annotators across diverse social and demographic backgrounds to try to understand whose values are represented. With ValuePrism, we build Kaleido, an open, light-weight, and structured language-based multi-task model that generates, explains, and assesses the relevance and valence (i.e., support or oppose) of human values, rights, and duties within a specific context. Humans prefer the sets of values output by our system over the teacher GPT-4, finding them more accurate and with broader coverage. In addition, we demonstrate that Kaleido can help explain variability in human decision-making by outputting contrasting values. Finally, we show that Kaleido's representations transfer to other philosophical frameworks and datasets, confirming the benefit of an explicit, modular, and interpretable approach to value pluralism. We hope that our work will serve as a step to making more explicit the implicit values behind human decision-making and to steering AI systems to make decisions that are more in accordance with them.

Full-duplex spoken dialogue systems must distinguish interruptions that require yielding the floor from backchannels that permit continued speaking. Existing benchmarks typically evaluate events independently and may therefore reward fixed action preferences rather than context-sensitive decisions. We introduce ECHO, a paired diagnostic benchmark for Chinese full-duplex turn-taking. ECHO pairs examples with the same overlap transcript but contrasting preceding multi-turn dialogue contexts, with one requiring Yield and the other Keep. It additionally includes off-talk examples for diagnosing unnecessary yielding. We introduce pair accuracy, which requires correct decisions on both members of a pair and assigns no credit to constant-action policies. Experiments on multiple full-duplex systems show that most exhibit a pronounced bias toward Yield, performing substantially better on interruptions than on backchannels, while another system remains comparatively balanced. These findings demonstrate that interruption-only evaluation can overestimate practical turn-taking reliability. ECHO and its metadata will be publicly released.

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

As language models shift from single-shot answer generation toward multi-step reasoning that retrieves and consumes evidence mid-inference, evaluating the role of individual retrieved items becomes more important. Existing RAG evaluation typically targets final-answer quality, citation faithfulness, or answer-level attribution, but none of these directly targets the intervention-based, per-evidence-item utility view we study here. We introduce CUE-R, a lightweight intervention-based framework for measuring per-evidence-item operational utility in single-shot RAG using shallow observable retrieval-use traces. CUE-R perturbs individual evidence items via REMOVE, REPLACE, and DUPLICATE operators, then measures changes along three utility axes (correctness, proxy-based grounding faithfulness, and confidence error) plus a trace-divergence signal. We also outline an operational evidence-role taxonomy for interpreting intervention outcomes. Experiments on HotpotQA and 2WikiMultihopQA with Qwen-3 8B and GPT-5.2 reveal a consistent pattern: REMOVE and REPLACE substantially harm correctness and grounding while producing large trace shifts, whereas DUPLICATE is often answer-redundant yet not fully behaviorally neutral. A zero-retrieval control confirms that these effects arise from degradation of meaningful retrieval. A two-support ablation further shows that multi-hop evidence items can interact non-additively: removing both supports harms performance far more than either single removal. Our results suggest that answer-only evaluation misses important evidence effects and that intervention-based utility analysis is a practical complement for RAG evaluation.

Instruction-tuned language models achieve strong performance across a range of generation tasks, but have also recently been shown to exhibit verbalized overconfidence. In question answering, verbalized model overconfidence may be associated with the consistency of the generated supporting rationales. In this paper, we study whether corresponding changes in the lexical diversity of generated answer rationales accompany changes in model confidence induced by instruction tuning. We evaluate three matched base and instruction-tuned models across question-answering benchmarks and find that instruction tuning consistently alters answer confidence, despite limited changes in predictive accuracy and decreases in likelihood-based calibration. Secondly, we observe a non-uniform effect of instruction tuning on rationale diversity: cross-rationale diversity consistently decreases, whereas surface-level lexical diversity varies in both direction and magnitude across models and benchmarks. Finally, we find that these differences persist after controlling for answer selection and rationale length, confirming that confidence and rationale diversity capture distinct effects of instruction tuning.

The increased use of Large Language Models (LLMs) in sensitive domains leads to growing interest in how their confidence scores correspond to fairness and bias. This study examines the alignment between LLM-predicted confidence and human-annotated bias judgments. Focusing on gender bias, the research investigates probability confidence calibration in contexts involving gendered pronoun resolution. The goal is to evaluate if calibration metrics based on predicted confidence scores effectively capture fairness-related disparities in LLMs. The results show that, among the six state-of-the-art models, Gemma-2 demonstrates the worst calibration according to the gender bias benchmark. The primary contribution of this work is a fairness-aware evaluation of LLMs' confidence calibration, offering guidance for ethical deployment. In addition, we introduce a new calibration metric, Gender-ECE, designed to measure gender disparities in resolution tasks.

Recent work has shown that models trained to the same objective, and which achieve similar measures of accuracy on consistent test data, may nonetheless behave very differently on individual predictions. This inconsistency is undesirable in high-stakes contexts, such as medical diagnosis and finance. We show that this inconsistent behavior extends beyond predictions to feature attributions, which may likewise have negative implications for the intelligibility of a model, and one's ability to find recourse for subjects. We then introduce selective ensembles to mitigate such inconsistencies by applying hypothesis testing to the predictions of a set of models trained using randomly-selected starting conditions; importantly, selective ensembles can abstain in cases where a consistent outcome cannot be achieved up to a specified confidence level. We prove that that prediction disagreement between selective ensembles is bounded, and empirically demonstrate that selective ensembles achieve consistent predictions and feature attributions while maintaining low abstention rates. On several benchmark datasets, selective ensembles reach zero inconsistently predicted points, with abstention rates as low 1.5%.

Despite its crucial role in research experiments, code correctness is often presumed only on the basis of the perceived quality of results. This assumption comes with the risk of erroneous outcomes and potentially misleading findings. To address this issue, we posit that the current focus on reproducibility should go hand in hand with the emphasis on software quality. We present a case study in which we identify and fix three bugs in widely used implementations of the state-of-the-art Conformer architecture. Through experiments on speech recognition and translation in various languages, we demonstrate that the presence of bugs does not prevent the achievement of good and reproducible results, which however can lead to incorrect conclusions that potentially misguide future research. As a countermeasure, we propose a Code-quality Checklist and release pangoliNN, a library dedicated to testing neural models, with the goal of promoting coding best practices and improving research software quality within the NLP community.

Speech brain-computer interfaces (speech BCIs) translate neural activity into language, offering a path towards restoring speech for people with paralysis and, more broadly, enabling new forms of natural human-computer interaction. Despite this promise, the field lacks a common measure of progress because systems use different datasets, recording methods, types of speech, and vocabularies, so their reported scores are rarely comparable. Underlying this measurement problem are two unresolved questions: (i) what distribution of words should a speech BCI enable a user to communicate, and (ii) how much information from this distribution can a system convey. We address both by deriving open-vocabulary mutual information (OVMI), an information-theoretic quantity that measures the information conveyed by a decoder relative to a reference distribution over the words a user may wish to communicate. This allows capabilities measured under different conditions, such as distinct vocabularies, to be evaluated on a common communication scale. We show that ordinarily reported accuracy, word error rate (WER), and other metrics computed only over the words a system supports can overstate how much of a user's intended speech the system can communicate. We then use OVMI to compare existing systems, expose trade-offs between how much of the user's language a system supports and how accurately it decodes those words, show that these comparisons depend on what the user is expected to communicate, and demonstrate that selecting a vocabulary to maximise OVMI yields up to 16.3% relative improvement in accuracy across three speech domains. OVMI therefore provides the speech BCI community with a principled way to compare heterogeneous systems, improve vocabulary design, and measure progress in the field.

We report a multi-month engineering program to build Sophea, a production bilingual Greek-English automatic speech recognition system. We evaluate the system against nine production gates covering Greek and English word error rate, language identification, and hallucinations on non-speech audio. Across twenty-three training iterations and two model architectures, no training-data composition passed all nine gates simultaneously. Meeting the Greek noisy-environment target required about 1,500 steps of dense domain exposure, while preserving English language identification tolerated only about 250 steps, or about 1,250 with a rebalanced mix that reduced Greek accuracy. We describe a six-stage data pipeline in which calibrating an audio-quality filter against in-domain anchors reduced the discarded share of scored Greek audio from 98.7 percent to 10.6 percent. A pre-registered ablation isolated a hallucination defect to one training-data package. A three-model ROVER ensemble increased gate coverage from 4-7 of 9 for individual models to 9 of 9 and reduced overlapping-speech WER from 53.35 percent to 37.87 percent, a 29 percent relative improvement. A separate learned per-clip arbiter over two models is listed as sophea/asr-k1 (preview) on the public Open ASR Leaderboard, with 4.26 percent average WER across eight public English test sets, and reaches 25.88 percent WER on live Greek noisy-environment traffic. We also document five cases in which a measurement tool produced a plausible but incorrect result and seven substantial approaches that were evaluated but not shipped. No model weights or training data are released; we report methodology and quantitative results only.

Knowing domain rules does not guarantee applying them to a case. We study this distinction in traditional Chinese Bazi through 3,000 Chinese multiple-choice questions spanning 14 Theory and 11 Case categories. Six endpoint systems are evaluated, with primary results reported on a 2,492-item model-informed refinement. Theory accuracy exceeds Case accuracy for every system, and gaps of 16.60-29.56 percentage points remain when invalid responses are excluded. The contrast is more specific than a general case-reasoning deficit. Across six systems, Twelve Stages and Nayin reach mean accuracies of 89.10% and 88.62%, while Shensha Basics reaches 75.96%. Within Case, Luck Pillars averages 84.62%, but Career and Family Relations average only 36.98% and 38.19%. Overall rankings also conceal different category strengths. On the original 3,000 items, paired DeepSeek native/disabled comparisons associate native configurations with Theory gains of 6.53 and 12.20 points for Flash and Pro, respectively; Case changes are -3.67 and +1.27 points. These are provider-configuration associations, not isolated causal effects of reasoning. The results motivate task-specific evaluation of cultural-domain applications rather than reliance on aggregate knowledge scores. The benchmark measures agreement with a model-generated, model-verified answer key, not real-world predictive validity. Final-set results are post-selection descriptions, and incomplete provenance and expert validation constrain their interpretation.

Speech and audio LLMs are evaluated by comparing waveform predictions with predictions from an automatic speech recognition (ASR) transcript. For fixed closed-set tasks, this conflates acoustic evidence with the need to invoke a generative audio model. We estimate incremental call value with matched selectors sharing pre-call evidence. Each policy may retain the transcript label, use a local encoder, or invoke a generative model; matched control removes generative actions but preserves pre-call evidence and development selection. On VocalSound, transcript-only accuracy is 0.296, while supervised CLAP and WavLM controls reach 0.850 and 0.854 without calls. Full selector reaches 0.925 at 12.5% calls versus 0.921 for matched No-call selector (difference 0.004; 95% CI [-0.025, 0.033]). Thus, results do not show a call gain after transcript and encoder evidence are available. Relevant quantity is incremental accuracy from allowing calls, not the waveform-transcript gap.

Assessing the fidelity and diversity of the generative model is a difficult but important issue for technological advancement. So, recent papers have introduced k-Nearest Neighbor (kNN) based precision-recall metrics to break down the statistical distance into fidelity and diversity. While they provide an intuitive method, we thoroughly analyze these metrics and identify oversimplified assumptions and undesirable properties of kNN that result in unreliable evaluation, such as susceptibility to outliers and insensitivity to distributional changes. Thus, we propose novel metrics, P-precision and P-recall (PP\&PR), based on a probabilistic approach that address the problems. Through extensive investigations on toy experiments and state-of-the-art generative models, we show that our PP\&PR provide more reliable estimates for comparing fidelity and diversity than the existing metrics. The codes are available at https://github.com/kdst-team/Probablistic_precision_recall.

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

Indic quality estimation (QE) and automatic post-editing (APE) data is spread across separate releases, so no single resource supports training and evaluation across tasks and language pairs on one footing. We consolidate the WMT 2020--2024 shared-task lineage with an extended English--Malayalam resource into \indicqe: 126{,}754 instances over nine directional pairs, with up to four label types aligned on the same segment, a direct assessment, a human post-edit, word-level OK/BAD tags and an error explanation, and a test set stratified over four difficulty axes. On it, we benchmark six prompted LLMs and three COMET metrics on segment-level QE, and three systems on APE. Two of the axes are defined partly on the direct assessment and select a compressed slice of it, so each axis is compared against a control drawn from the same language pair with the same score distribution. Only one survives that control: segments whose holistic and token-level quality signals conflict are ranked worse than equally-scored segments of the same language, for all nine systems and all seven pairs that carry the axis. Annotator disagreement, which looks second-hardest without the control, has no effect with it. Few-shot prompting costs every model leq 3.4B both correlation and output-format compliance. Within-language accuracy does not make scores comparable across pairs: of the three trained metrics, the one with the best within-language correlation loses most when the pairs are pooled. The benchmark and code will be released.

Prediction problems often admit competing models that perform almost equally well. This effect challenges key assumptions in machine learning when competing models assign conflicting predictions. In this paper, we define predictive multiplicity as the ability of a prediction problem to admit competing models with conflicting predictions. We introduce formal measures to evaluate the severity of predictive multiplicity and develop integer programming tools to compute them exactly for linear classification problems. We apply our tools to measure predictive multiplicity in recidivism prediction problems. Our results show that real-world datasets may admit competing models that assign wildly conflicting predictions, and motivate the need to measure and report predictive multiplicity in model development.

Recent advancements in large language models (LLMs) have demonstrated remarkable reasoning capabilities. However, single-shot inference often yields unreliable results for complex reasoning tasks, leading researchers to explore multiple reasoning paths through methods such as perplexity and self-consistency. In this paper, we present the first theoretical error decomposition analysis of these techniques, breaking down their error into estimation error and model error. Our analysis reveals a fundamental trade-off: perplexity methods suffer from substantial model error due to the absence of a proper consistency function, while self-consistency exhibits high estimation error due to a slow error convergence rate. To overcome these limitations, we propose Reasoning-Pruning Perplexity Consistency (RPC). This approach combines Perplexity Consistency, which seamlessly integrates LLM perplexity with self-consistency, and Reasoning Pruning, which eliminates low-probability reasoning paths to effectively prevent the degeneration of estimation error reduction. Theoretical analysis demonstrates that RPC not only accelerates the convergence rate of estimation error to an exponential level but also holds strong potential for further reducing model error. Extensive empirical evaluations on seven benchmark datasets confirm that RPC can significantly improve reasoning performance, sample efficiency, and confidence reliability.

We evaluated 4 systems (ELIZA, GPT-4o, LLaMa-3.1-405B, and GPT-4.5) in two randomised, controlled, and pre-registered Turing tests on independent populations. Participants had 5 minute conversations simultaneously with another human participant and one of these systems before judging which conversational partner they thought was human. When prompted to adopt a humanlike persona, GPT-4.5 was judged to be the human 73% of the time: significantly more often than interrogators selected the real human participant. LLaMa-3.1, with the same prompt, was judged to be the human 56% of the time -- not significantly more or less often than the humans they were being compared to -- while baseline models (ELIZA and GPT-4o) achieved win rates significantly below chance (23% and 21% respectively). The results constitute the first empirical evidence that any artificial system passes a standard three-party Turing test. The results have implications for debates about what kind of intelligence is exhibited by Large Language Models (LLMs), and the social and economic impacts these systems are likely to have.

Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70\% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

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

Self-improving LLM systems propose changes to themselves and keep those that score better on a small evaluation set. We treat this keep-if-better step as selection under measurement noise, model the correlated errors of the candidates in a single decision, and study empirically what happens when the evaluation set is reused. In runs where Qwen models rewrite their own instructions and every candidate is also scored on 600 held-out items, most proposals after the first are harmful, and the model gives the size of the winner's curse of a generation's best candidate. With a prior from a separate pilot, it matches the average overstatement of first-generation commits in native loops, though not setting by setting. In a pre-registered study, the final selection-set score of greedy loops exceeded held-out accuracy by 13 to 20 points with 16 selection items and by 1 to 5 points with 256. Held-out gains grew with the selection set on TREC but not on GSM8K, and the tested acceptance rules did not beat greedy acceptance over whole runs. Gains measured on the selection set also exceeded held-out gains when a current model refined a competent instruction, and in the validation scores of GEPA and MIPROv2. Scoring the starting and the current instruction on 64 items never used for selection removes the average bias of a loop's reported gain, but single estimates remain off by about 6 points. Self-improvement studies should report held-out gains with their uncertainty.

Language models (LMs) should provide reliable confidence estimates to help users detect mistakes in their outputs and defer to human experts when necessary. Asking a language model to assess its confidence ("Score your confidence from 0-1.") is a natural way of evaluating its uncertainty. However, models struggle to provide absolute assessments of confidence (i.e. judging confidence in answering a question independent of other questions) and the coarse-grained scores they produce are not useful for evaluating the correctness of their answers. We propose relative confidence estimation, where we match up questions against each other and ask the model to make relative judgments of confidence ("Which question are you more confident in answering correctly?"). Treating each question as a "player" in a series of matchups against other questions and the model's preferences as match outcomes, we can use rank aggregation methods like Elo rating and Bradley-Terry to translate the model's confidence preferences into confidence scores. We evaluate relative confidence estimation against absolute confidence estimation and self-consistency confidence methods on five state-of-the-art LMs -- GPT-4, GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet, and Llama 3.1 405B -- across 14 challenging STEM, social science, and commonsense reasoning question answering tasks. Our results demonstrate that relative confidence estimation consistently provides more reliable confidence scores than absolute confidence estimation, with average gains of 3.5% in selective classification AUC over direct absolute confidence estimation methods and 1.7% over self-consistency approaches across all models and datasets.

We present "Human or Not?", an online game inspired by the Turing test, that measures the capability of AI chatbots to mimic humans in dialog, and of humans to tell bots from other humans. Over the course of a month, the game was played by over 1.5 million users who engaged in anonymous two-minute chat sessions with either another human or an AI language model which was prompted to behave like humans. The task of the players was to correctly guess whether they spoke to a person or to an AI. This largest scale Turing-style test conducted to date revealed some interesting facts. For example, overall users guessed the identity of their partners correctly in only 68% of the games. In the subset of the games in which users faced an AI bot, users had even lower correct guess rates of 60% (that is, not much higher than chance). This white paper details the development, deployment, and results of this unique experiment. While this experiment calls for many extensions and refinements, these findings already begin to shed light on the inevitable near future which will commingle humans and AI.

In industry NLP application, our manually labeled data has a certain number of noisy data. We present a simple method to find the noisy data and relabel them manually, meanwhile we collect the correction information. Then we present novel method to incorporate the human correction information into deep learning model. Human know how to correct noisy data. So the correction information can be inject into deep learning model. We do the experiment on our own text classification dataset, which is manually labeled, because we need to relabel the noisy data in our dataset for our industry application. The experiment result shows that our learn-on-correction method improve the classification accuracy from 91.7% to 92.5% in test dataset. The 91.7% accuracy is trained on the corrected dataset, which improve the baseline from 83.3% to 91.7% in test dataset. The accuracy under human evaluation achieves more than 97%.

Combating bias in NLP requires bias measurement. Bias measurement is almost always achieved by using lexicons of seed terms, i.e. sets of words specifying stereotypes or dimensions of interest. This reproducibility study focuses on the original authors' main claim that the rationale for the construction of these lexicons needs thorough checking before usage, as the seeds used for bias measurement can themselves exhibit biases. The study aims to evaluate the reproducibility of the quantitative and qualitative results presented in the paper and the conclusions drawn thereof. We reproduce most of the results supporting the original authors' general claim: seed sets often suffer from biases that affect their performance as a baseline for bias metrics. Generally, our results mirror the original paper's. They are slightly different on select occasions, but not in ways that undermine the paper's general intent to show the fragility of seed sets.

Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks' CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.

We test whether Large Language Models (LLMs) can be used to simulate human participants in social-science studies. To do this, we run replications of 14 studies from the Many Labs 2 replication project with OpenAI's text-davinci-003 model, colloquially known as GPT3.5. Based on our pre-registered analyses, we find that among the eight studies we could analyse, our GPT sample replicated 37.5% of the original results and 37.5% of the Many Labs 2 results. However, we were unable to analyse the remaining six studies due to an unexpected phenomenon we call the "correct answer" effect. Different runs of GPT3.5 answered nuanced questions probing political orientation, economic preference, judgement, and moral philosophy with zero or near-zero variation in responses: with the supposedly "correct answer." In one exploratory follow-up study, we found that a "correct answer" was robust to changing the demographic details that precede the prompt. In another, we found that most but not all "correct answers" were robust to changing the order of answer choices. One of our most striking findings occurred in our replication of the Moral Foundations Theory survey results, where we found GPT3.5 identifying as a political conservative in 99.6% of the cases, and as a liberal in 99.3% of the cases in the reverse-order condition. However, both self-reported 'GPT conservatives' and 'GPT liberals' showed right-leaning moral foundations. Our results cast doubts on the validity of using LLMs as a general replacement for human participants in the social sciences. Our results also raise concerns that a hypothetical AI-led future may be subject to a diminished diversity-of-thought.

Full-duplex evaluation often emphasizes whether an agent keeps speaking or stops. That binary cannot express a third response humans use routinely: continuing to speak while incorporating what the listener just contributed. The contribution may be a missing word, a correction or a clarification. We introduce Duplex Cue, an evaluation of this in-turn adaptation in full-duplex voice agents. Duplex Cue separates listener intent (backchannel, collaboration, or interruption) from speaker behavior: continuing unchanged, adapting within the turn, or yielding. Adaptation includes acknowledgment as well as content revision. In a single-model case study using 300 human-confirmed cues from unscripted English conversations, we compare recorded human responses with PersonaPlex continuations generated while replaying the listener's audio. We retain 208 pairs with the ongoing speaker active at cue onset and a scorable response in each condition. On the 66 collaborative pairs, recorded speakers adapt in 68.2\% of cases, compared with 34.8\% for PersonaPlex. The model otherwise continues unchanged (42.4\%) or yields (22.7\%). These findings show why evaluating natural voice interaction requires measuring how an agent responds to a listener's contribution as well as whether it keeps speaking.

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

Large Language Models often improve accuracy on reasoning tasks by sampling multiple Chain-of-Thought (CoT) traces and aggregating them with majority voting (MV), a test-time technique called self-consistency. When we truncate a CoT partway through and regenerate the remainder, we observe that traces with correct answers reproduce their original answer more often than traces with wrong answers. We use this difference as a reliability signal, prefix consistency, that weights each candidate answer by how often it reappears under regeneration. It requires no access to token log-probabilities or self-rating prompts. Across five reasoning models and four math and science benchmarks, prefix consistency is the best correctness predictor in most settings, and reweighting votes by it reaches Standard MV plateau accuracy at up to 21x fewer tokens (median 4.6x). Our code is available at https://github.com/naoto-iwase/prefix-consistency.

An algorithm to estimate the evolution of learning curves on the whole of a training data base, based on the results obtained from a portion and using a functional strategy, is introduced. We approximate iteratively the sought value at the desired time, independently of the learning technique used and once a point in the process, called prediction level, has been passed. The proposal proves to be formally correct with respect to our working hypotheses and includes a reliable proximity condition. This allows the user to fix a convergence threshold with respect to the accuracy finally achievable, which extends the concept of stopping criterion and seems to be effective even in the presence of distorting observations. Our aim is to evaluate the training effort, supporting decision making in order to reduce the need for both human and computational resources during the learning process. The proposal is of interest in at least three operational procedures. The first is the anticipation of accuracy gain, with the purpose of measuring how much work is needed to achieve a certain degree of performance. The second relates the comparison of efficiency between systems at training time, with the objective of completing this task only for the one that best suits our requirements. The prediction of accuracy is also a valuable item of information for customizing systems, since we can estimate in advance the impact of settings on both the performance and the development costs. Using the generation of part-of-speech taggers as an example application, the experimental results are consistent with our expectations.

Quality estimation models have been developed to assess the corrections made by grammatical error correction (GEC) models when the reference or gold-standard corrections are not available. An ideal quality estimator can be utilized to combine the outputs of multiple GEC systems by choosing the best subset of edits from the union of all edits proposed by the GEC base systems. However, we found that existing GEC quality estimation models are not good enough in differentiating good corrections from bad ones, resulting in a low F0.5 score when used for system combination. In this paper, we propose GRECO, a new state-of-the-art quality estimation model that gives a better estimate of the quality of a corrected sentence, as indicated by having a higher correlation to the F0.5 score of a corrected sentence. It results in a combined GEC system with a higher F0.5 score. We also propose three methods for utilizing GEC quality estimation models for system combination with varying generality: model-agnostic, model-agnostic with voting bias, and model-dependent method. The combined GEC system outperforms the state of the art on the CoNLL-2014 test set and the BEA-2019 test set, achieving the highest F0.5 scores published to date.

Large Language Models (LLMs) can correct their self-generated responses, but a decline in accuracy after self-correction is also witnessed. To have a deeper understanding of self-correction, we endeavor to decompose, evaluate, and analyze the self-correction behaviors of LLMs. By enumerating and analyzing answer correctness before and after self-correction, we decompose the self-correction capability into confidence (being confident to correct answers) and critique (turning wrong answers to correct) capabilities, and propose two metrics from a probabilistic perspective to measure these 2 capabilities, along with another metric for overall self-correction capability evaluation. Based on our decomposition and evaluation metrics, we conduct extensive experiments and draw some empirical conclusions. For example, we find different models can exhibit distinct behaviors: some models are confident while others are more critical. We also find the trade-off between the two capabilities (i.e. improving one can lead to a decline in the other) when manipulating model self-correction behavior by prompts or in-context learning. Further, we find a simple yet efficient strategy to improve self-correction capability by transforming Supervision Fine-Tuning (SFT) data format, and our strategy outperforms vanilla SFT in both capabilities and achieves much higher accuracy after self-correction. Our code will be publicly available on GitHub.

We introduce Full-Duplex-Bench-v3 (FDB-v3), a benchmark for evaluating spoken language models under naturalistic speech conditions and multi-step tool use. Unlike prior work, our dataset consists entirely of real human audio annotated for five disfluency categories, paired with scenarios requiring chained API calls across four task domains. We evaluate six model configurations -- GPT-Realtime, Gemini Live 2.5, Gemini Live 3.1, Grok, Ultravox v0.7, and a traditional Cascaded pipeline (WhisperrightarrowGPT-4orightarrowTTS) -- across accuracy, latency, and turn-taking dimensions. GPT-Realtime leads on Pass@1 (0.600) and interruption avoidance (13.5\%); Gemini Live 3.1 achieves the fastest latency (4.25~s) but the lowest turn-take rate (78.0\%); and the Cascaded baseline, despite a perfect turn-take rate, incurs the highest latency (10.12~s). Across all systems, self-correction handling and multi-step reasoning under hard scenarios remain the most consistent failure modes.

Some security-development settings require local models that map an objective to an ordered, checkable plan. We present a low-rank decomposition adapter and release a case-disjoint corpus with 24 authored decompositions and 83 derived next-step examples across 24 satellite-security cases. To prevent reference-plan leakage, prompts contain one objective and eight shuffled SPARTA candidates. The evaluation candidate sets are constructed with oracle inclusion of every reference technique, so this study measures selection from a controlled set, not retrieval. On six fixed cases and five decoding seeds, mean scores vary across adapters and prompted Qwen2.5 baselines; a separate greedy check over three adapter training seeds shows non-negligible training variance. A 24-fold greedy leave-one-case-out control at the two compact sizes likewise shows tradeoffs: at 1.5B, adapter precision is 0.583 against 0.480 for two-shot prompting, while recall is 0.586 against 0.660. Case-resampling intervals for adapter precision relative to the strongest prompted baseline cross zero at 1.5B and 7B. The adapter is also substantially more format-compliant, which prevents attributing all score differences to selection. Autoregressive next-step accuracy is 0.08, 0.06, and 0.29 at 0.5B, 1.5B, and 7B. We therefore offer the formulation, leakage-controlled dataset, reproducibility package, and descriptive analysis as a proof of concept, not evidence of a reliable stand-alone decomposition system or a general adapter advantage. A separate diagnostic tests contract reuse.

The increasing reliance on AI-driven solutions, particularly Large Language Models (LLMs) like the GPT series, for information retrieval highlights the critical need for their factuality and fairness, especially amidst the rampant spread of misinformation and disinformation online. Our study evaluates the factual accuracy, stability, and biases in widely adopted GPT models, including GPT-3.5 and GPT-4, contributing to reliability and integrity of AI-mediated information dissemination. We introduce 'Global-Liar,' a dataset uniquely balanced in terms of geographic and temporal representation, facilitating a more nuanced evaluation of LLM biases. Our analysis reveals that newer iterations of GPT models do not always equate to improved performance. Notably, the GPT-4 version from March demonstrates higher factual accuracy than its subsequent June release. Furthermore, a concerning bias is observed, privileging statements from the Global North over the Global South, thus potentially exacerbating existing informational inequities. Regions such as Africa and the Middle East are at a disadvantage, with much lower factual accuracy. The performance fluctuations over time suggest that model updates may not consistently benefit all regions equally. Our study also offers insights into the impact of various LLM configuration settings, such as binary decision forcing, model re-runs and temperature, on model's factuality. Models constrained to binary (true/false) choices exhibit reduced factuality compared to those allowing an 'unclear' option. Single inference at a low temperature setting matches the reliability of majority voting across various configurations. The insights gained highlight the need for culturally diverse and geographically inclusive model training and evaluation. This approach is key to achieving global equity in technology, distributing AI benefits fairly worldwide.

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

Decision models, non-generative scorers that report a probability over caller-supplied options and emit no text, are increasingly deployed as routers, classifiers and guardrails. Because they generate nothing, rationales and chain-of-thought do not apply. We audit three: an encoder scorer, a decoder scorer and a closed API, entirely through interventions on their input, using expert-annotated Bengali idioms carrying a gold figurative and a gold literal gloss. Dropping an idiom's own literal gloss into the option list draws every arm toward it, taking 65% to 96% of each arm's remaining Bengali errors. A model-by-language matrix over one shared item set shows how far this depends on the test language: every arm is 1.9 to 6.4 times more captured in Bengali than in English on the same paired entries, the commercial model's apparent resistance is a property of its test language, and capability does not order the failure. Calibration degrades in all three arms, and on the open arms confidence inverts, so gating on it is worse than gating at random. On a semantics-free option list, where accuracy is at chance, the best-calibrated arm still reports mean confidence 0.82. Code and data: https://github.com/sakhadib/literal-lock-in.

Language models are appealing tools for research on the past. But to trust the evidence a model provides, researchers need to know whether its responses fit the period represented. Validation is challenging, because this is not a task living people ordinarily perform, and because many questions have multiple correct answers. We use historical texts to develop a benchmark for a model's representation of English-language contexts 1831-1930, relying on pairwise comparisons to multiple ground truths and strong distractors to score the hardest questions in an appropriately graduated way. We find that generative tasks are harder than discriminative ones; in fact, reasoning models can typically discern the weakness of their own generated answers. While models pretrained exclusively on historical text lead the pack when evaluated by answer likelihood, they cannot compete with commercial models in free generation. None of the models we tested represent historical contexts in a fully persuasive way yet, but progress toward that goal is evident.

Decision models support routing, guardrails, and output verification in LLM systems. However, item accuracy alone cannot establish whether models make correct decisions across controlled variations or assign reliable confidence to their predictions. We introduce DecisionPivot, a diagnostic benchmark comprising 48 tasks and 40,033 items across seven decision roles. Its core track contains 7,800 items from 13 generated task families, organized into contrast groups with computed answers verified by independently implemented solvers. The complementary reference track contains 32,233 items from 35 public test sets, preserving their original evaluation metrics. We evaluate 40 decision models and 14 general models using documented, model-specific protocols, assessing item and contrast-group accuracy, probability ordering, calibration, and empirical selective coverage. Our results reveal substantial differences obscured by aggregate scores: two 27B models separated by only 2.6 points on the reference track achieve contrast-group accuracies of 70.1% and 44.2%. Across 2,400 binary contrast pairs, Perplexity Decider correctly ranks the positive member above the negative member in 92.3% of pairs but answers both members correctly in only 74.1%, exposing a gap between probability ordering and decision correctness. Models with similar item accuracy also differ in the reliability of their high-confidence predictions. On a separate 1,296-item subset, GLM-5.3 achieves 93.5\% contrast-group accuracy at a higher reasoning setting, compared with 81.5% at its lowest effort. By linking errors to pivotal details and difficulty levels, DecisionPivot provides fine-grained diagnostics of decision failures that aggregate evaluation alone cannot reveal.

How fast does a language model degrade when trained on its own outputs? Theory traces it to gradually accumulating errors, while experiments report repeated phrases within ten generations. Under a fixed sampling seed in vLLM, the fast loss of lexical diversity comes from the sampler. When vLLM serves a batch from one seeded sampling configuration, every request receives the same random draws, and a fixed seed replays them every generation. Fine-tuning raises the tokens that won, and the replayed draws let them win by more. Sharing across requests and replay across generations matter only together. Remove either one, by changing the shared seed every generation or by giving each request its own seed that repeats every generation, and the unique-4-gram fraction of two StableLM checkpoints stays near its starting value of about 0.98 through generation 3. Keep both, and the replayed shared seed takes seven checkpoints from five families to between 0.045 and 0.38 by then. Three generations of replay write the favoured phrases into the weights: decoded with one seed per request, the generation-3 weights of the replayed StableLM-2-1.6B chain recover most of their diversity, yet the phrase that filled every sample under the shared seed still opens 46% of them. Without replay, five checkpoints drift slowly, consistent with the gradual accumulation that theory describes, and three turn incoherent though their diversity scores stay high. One peer-reviewed model-collapse pipeline that fine-tunes Gemma-2-27B samples identical prompts under one seeded configuration, and three quarters of the rows it released for one iteration repeat nearly as often as one such batch copies them. A seed per request restores the fresh sample that stability analyses assume.

Latent communication passes internal states between language models instead of decoded text, but higher receiver accuracy does not show that the receiver used the message content. Across five method-dataset pairs, replacing each message with one from an unrelated question changes accuracy by at most 0.60 points, even when communication adds 15.44 points over the receiver alone. Thus the interface can supply the gain while making the sharer dispensable. Draft-KV instead sends the key-value states formed while the sharer drafts an answer to the current question. Linear projections place these states in a side memory read through a gated attention branch, and progressive training moves from message reconstruction to answer supervision under a guard on harm from mismatched messages. Both models remain frozen and the interface trains 1.05M parameters, 348x fewer than C2C. With a Qwen3-8B sharer, a frozen Qwen2.5-0.5B-Instruct receiver reaches 78.04% on MMLU-Redux, versus 37.45% alone and 36.40% with reassigned messages. At fixed interface size, scaling the sharer from 0.6B to 8B raises accuracy from 46.11% to 78.04%; communication also transfers to held-out tasks and can exceed both models when each holds different evidence.

While existing alignment paradigms have been integral in developing large language models (LLMs), LLMs often learn an averaged human preference and struggle to model diverse preferences across cultures, demographics, and communities. We propose Modular Pluralism, a modular framework based on multi-LLM collaboration for pluralistic alignment: it "plugs into" a base LLM a pool of smaller but specialized community LMs, where models collaborate in distinct modes to flexibility support three modes of pluralism: Overton, steerable, and distributional. Modular Pluralism is uniquely compatible with black-box LLMs and offers the modular control of adding new community LMs for previously underrepresented communities. We evaluate Modular Pluralism with six tasks and four datasets featuring questions/instructions with value-laden and perspective-informed responses. Extensive experiments demonstrate that Modular Pluralism advances the three pluralism objectives across six black-box and open-source LLMs. Further analysis reveals that LLMs are generally faithful to the inputs from smaller community LLMs, allowing seamless patching by adding a new community LM to better cover previously underrepresented communities.

Language models have demonstrated remarkable performance in solving reasoning tasks; however, even the strongest models still occasionally make reasoning mistakes. Recently, there has been active research aimed at improving reasoning accuracy, particularly by using pretrained language models to "self-correct" their mistakes via multi-round prompting. In this paper, we follow this line of work but focus on understanding the usefulness of incorporating "error-correction" data directly into the pretraining stage. This data consists of erroneous solution steps immediately followed by their corrections. Using a synthetic math dataset, we show promising results: this type of pretrain data can help language models achieve higher reasoning accuracy directly (i.e., through simple auto-regression, without multi-round prompting) compared to pretraining on the same amount of error-free data. We also delve into many details, such as (1) how this approach differs from beam search, (2) how such data can be prepared, (3) whether masking is needed on the erroneous tokens, (4) the amount of error required, (5) whether such data can be deferred to the fine-tuning stage, and many others.

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are executed. The same candidate budget can be produced in one batched generation call or split across several sequential calls with smaller batch sizes. We first study the effect of increasing N on reasoning accuracy using Phi-3-mini and Qwen2.5-1.5B on 500 GSM8K prompts. As expected, increasing N from 1 to 8 improves accuracy by 8.4 percentage points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B. However, accuracy alone does not show the systems cost of using a larger candidate budget. We therefore fix N = 8 and compare four generation schedules: 1x8, 2x4, 4x2, and 8x1, where axb denotes a generation calls with b candidates per call. We measure latency, throughput, GPU-hours, and gross GPU-device energy while keeping the total candidate count fixed. On A100 GPUs, eight serial calls use 4.64-4.86x as much gross GPU-device energy and have 5.77-6.12x the P95 latency of one batched call with eight candidates. The same pattern appears across three independently scheduled A100 nodes per model and in short-output SciQ/V100 experiments. These results show that candidate count alone is not enough to describe the systems cost of multi-candidate test-time scaling. When candidates are independent and memory allows it, fewer generation calls with larger batch sizes are more efficient. Evaluations should therefore report not only candidate count and accuracy, but also generation schedule and GPU-level systems metrics.

Domestic AI agents faces ethical, autonomy, and inclusion challenges, particularly for overlooked groups like children, elderly, and Neurodivergent users. We present the Plural Voices Model (PVM), a novel single-agent framework that dynamically negotiates multi-user needs through real-time value alignment, leveraging diverse public datasets on mental health, eldercare, education, and moral reasoning. Using human+synthetic curriculum design with fairness-aware scenarios and ethical enhancements, PVM identifies core values, conflicts, and accessibility requirements to inform inclusive principles. Our privacy-focused prototype features adaptive safety scaffolds, tailored interactions (e.g., step-by-step guidance for Neurodivergent users, simple wording for children), and equitable conflict resolution. In preliminary evaluations, PVM outperforms multi-agent baselines in compliance (76% vs. 70%), fairness (90% vs. 85%), safety-violation rate (0% vs. 7%), and latency. Design innovations, including video guidance, autonomy sliders, family hubs, and adaptive safety dashboards, demonstrate new directions for ethical and inclusive domestic AI, for building user-centered agentic systems in plural domestic contexts. Our Codes and Model are been open sourced, available for reproduction: https://github.com/zade90/Agora

Verification for retrieval-augmented generation usually scores each retrieved chunk and drops the ones that fail. We show this cannot work for multi-hop questions, and show what does. Per-chunk scoring assumes one chunk is a sufficient premise for the answer. Multi-hop questions are built so that none is, and the paragraph carrying the answer is the one the question does not name. Entailment scoring reaches 0.643, 0.523 and 0.560 AUC on HotpotQA, 2WikiMultihopQA and MuSiQue, against 0.951 on single-hop SQuAD. Seven controls rule out model capacity, premise length, hypothesis template, decision threshold, retriever, answer-matching criterion and prompt. End to end across three datasets, three generator sizes and two prompts, per-chunk gating is significantly worse than not filtering at all in every cell, and its penalty grows with generator capability. The repair is to condition verification on the decomposed sub-question rather than the original query. Using MuSiQue's gold decomposition, entailment on a later hop rises from 0.546, which is chance, to 0.840, a paired lift of +0.355 with a bootstrap interval of [0.331, 0.382]. An off-the-shelf Qwen2.5-7B decomposer, given the question and the top retrieved paragraph, reaches 0.637 and captures 31% of that ceiling; decomposing without retrieval reaches 0.533, below the original question. Iterative retrieval systems already produce such decompositions and discard them before verifying.

As Large Language Models (LLMs) are nondeterministic, the same input can generate different outputs, some of which may be incorrect or hallucinated. If run again, the LLM may correct itself and produce the correct answer. Unfortunately, most LLM-powered systems resort to single results which, correct or not, users accept. Having the LLM produce multiple outputs may help identify disagreements or alternatives. However, it is not obvious how the user will interpret conflicts or inconsistencies. To this end, we investigate how users perceive the AI model and comprehend the generated information when they receive multiple, potentially inconsistent, outputs. Through a preliminary study, we identified five types of output inconsistencies. Based on these categories, we conducted a study (N=252) in which participants were given one or more LLM-generated passages to an information-seeking question. We found that inconsistency within multiple LLM-generated outputs lowered the participants' perceived AI capacity, while also increasing their comprehension of the given information. Specifically, we observed that this positive effect of inconsistencies was most significant for participants who read two passages, compared to those who read three. Based on these findings, we present design implications that, instead of regarding LLM output inconsistencies as a drawback, we can reveal the potential inconsistencies to transparently indicate the limitations of these models and promote critical LLM usage.

Large language models (LLMs) often produce errors, including factual inaccuracies, biases, and reasoning failures, collectively referred to as "hallucinations". Recent studies have demonstrated that LLMs' internal states encode information regarding the truthfulness of their outputs, and that this information can be utilized to detect errors. In this work, we show that the internal representations of LLMs encode much more information about truthfulness than previously recognized. We first discover that the truthfulness information is concentrated in specific tokens, and leveraging this property significantly enhances error detection performance. Yet, we show that such error detectors fail to generalize across datasets, implying that -- contrary to prior claims -- truthfulness encoding is not universal but rather multifaceted. Next, we show that internal representations can also be used for predicting the types of errors the model is likely to make, facilitating the development of tailored mitigation strategies. Lastly, we reveal a discrepancy between LLMs' internal encoding and external behavior: they may encode the correct answer, yet consistently generate an incorrect one. Taken together, these insights deepen our understanding of LLM errors from the model's internal perspective, which can guide future research on enhancing error analysis and mitigation.

Research on Large Language Models (LLMs) studies output variation across generation, reasoning, alignment, and representational analysis, often under the umbrella of "diversity." Yet the terminology remains fragmented, largely because the normative objectives underlying tasks are rarely made explicit. We introduce the Magic, Madness, Heaven, Sin framework, which models output variation along a homogeneity-heterogeneity axis, where valuation is determined by the task and its normative objective. We organize tasks into four normative contexts: epistemic (factuality), interactional (user utility), societal (representation), and safety (robustness). For each, we examine the failure modes and vocabulary such as hallucination, mode collapse, bias, and erasure through which variation is studied. We apply the framework to analyze all pairwise cross-contextual interactions, revealing that optimizing for one objective, such as improving safety, can inadvertently harm demographic representation or creative diversity. We argue for context-aware evaluation of output variation, reframing it as a property shaped by task objectives rather than a model's intrinsic trait.

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

Three of the most popular methods for training language models to reason look like three different tricks. They are not. All three adjust a single number: standard deviation, reflecting how much a prompt's sampled answers disagree. When such a model is trained, it answers each problem many times, and an automatic checker marks every answer right or wrong. The standard deviation of those marks measures the disagreement: largest when the answers split evenly between right and wrong, and zero when they all agree. Group Relative Policy Optimization (GRPO) divides by this number, GRPO Done Right (Dr. GRPO) drops the division, and Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) discards the groups where it is zero. Each is presented as its own fix, yet this paper proves they are three settings of one dial. That dial is not cosmetic: for right-or-wrong rewards, the disagreement is exactly the size of the training update, the group-standard-deviation identity. A split group teaches the most, while a unanimous group teaches nothing and falls silent. The same result says which problems deserve the most weight and how many tries each one needs. This paper confirms the intuition on a large real difficulty dataset (Big-Math) and in a controlled training run. What looks like a harmless normalization step is the dial that decides where learning happens and how strongly.

Accurate confidence estimation is essential for trustworthy large language models (LLMs) systems, as it empowers the user to determine when to trust outputs and enables reliable deployment in safety-critical applications. Current confidence estimation methods for LLMs neglect the relevance between responses and contextual information, a crucial factor in output quality evaluation, particularly in scenarios where background knowledge is provided. To bridge this gap, we propose CRUX (Context-aware entropy Reduction and Unified consistency eXamination), the first framework that integrates context faithfulness and consistency for confidence estimation via two novel metrics. First, contextual entropy reduction represents data uncertainty with the information gain through contrastive sampling with and without context. Second, unified consistency examination captures potential model uncertainty through the global consistency of the generated answers with and without context. Experiments across three benchmark datasets (CoQA, SQuAD, QuAC) and two domain-specific datasets (BioASQ, EduQG) demonstrate CRUX's effectiveness, achieving the highest AUROC than existing baselines.

Recent work has demonstrated the viability of using crowdsourcing as a tool for evaluating the truthfulness of public statements. Under certain conditions such as: (1) having a balanced set of workers with different backgrounds and cognitive abilities; (2) using an adequate set of mechanisms to control the quality of the collected data; and (3) using a coarse grained assessment scale, the crowd can provide reliable identification of fake news. However, fake news are a subtle matter: statements can be just biased ("cherrypicked"), imprecise, wrong, etc. and the unidimensional truth scale used in existing work cannot account for such differences. In this paper we propose a multidimensional notion of truthfulness and we ask the crowd workers to assess seven different dimensions of truthfulness selected based on existing literature: Correctness, Neutrality, Comprehensibility, Precision, Completeness, Speaker's Trustworthiness, and Informativeness. We deploy a set of quality control mechanisms to ensure that the thousands of assessments collected on 180 publicly available fact-checked statements distributed over two datasets are of adequate quality, including a custom search engine used by the crowd workers to find web pages supporting their truthfulness assessments. A comprehensive analysis of crowdsourced judgments shows that: (1) the crowdsourced assessments are reliable when compared to an expert-provided gold standard; (2) the proposed dimensions of truthfulness capture independent pieces of information; (3) the crowdsourcing task can be easily learned by the workers; and (4) the resulting assessments provide a useful basis for a more complete estimation of statement truthfulness.

Why do language models sometimes prefer correct statements even when trained on mixed-quality data? We introduce the Compression--Consistency Principle: next-token prediction favors hypotheses that allow shorter and more internally consistent descriptions of the training data. Truth bias emerges only when false alternatives are structurally harder to compress. We test this using small GPT-2-style character-level transformers (3.5M--86M parameters) on synthetic math corpora with controlled mixtures of correct and incorrect rules. In the random-error setting, models strongly prefer correct completions in paired evaluation: 83.1% accuracy at balanced data and 67.0% even when correct rules appear in only 10% of the corpus. Replacing random errors with a coherent but mathematically incorrect rule system largely eliminates the preference (near-chance accuracy). In a more natural-language-like synthetic world, the effect is weaker but still present (57.7%). Additional experiments show that embedding verification steps can restore preference for correctness even at small scale, while increasing the number of consistent rules produces a graded improvement in accuracy. Our results suggest that what appears as a "truth bias" is largely a side effect of compression pressure and preference for internal consistency, rather than an intrinsic drive toward truth. Full code and data are available at https://github.com/Rai220/compression-drives-truth.

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

Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar's tests for paired comparisons, and Cohen's d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.