Daily Papers

Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers. Under sustained adversarial pressure we find a previously undocumented failure mode: the chain-of-thought stays factually correct from first turn to last while the emitted answer flips wrong. We call this unfaithful capitulation (UC) and isolate it with a 2times 2 latent-versus-behavioral framework that flip-rate metrics and single-turn faithfulness probes both miss. Across three datasets (MT-Consistency, MMLU-Pro, GSM8K), the latent-correct rate at the behavioral flip clusters near 50% in think mode and collapses to 11-15% under no_think -- paired, within-model causal evidence that reasoning creates the gap. Across models the effect tracks the reasoning channel (high in Qwen3-32B and GPT-OSS-20B, low in inline-CoT Gemma-4-31B-it). An independent GPT-4o judge corroborates 86% of UC labels; a token-level probe shows the answer-slot argmax is correct in 84% of UC cells; and a naive trace-anchored defense backfires. We release all trajectories, traces, and judge labels.

Large language models often exhibit sycophancy, revising their answers to align with users when users push back. Such answer flips, however, can arise from different causes. One possibility is that the model simply aligns with the user's feedback in order to satisfy them. Another is that the feedback genuinely contains useful evidence, prompting the model to update its answer in a rational way. We distinguish them as Unsupported-Yielding and Rational-Updating. Prior work focuses primarily on suppressing Unsupported-Yielding, while overlooking its effect on Rational-Updating. We address this gap with a two-turn evaluation framework that measures the two behaviors separately. Across representative training-time and inference-time interventions, we find that anti-sycophancy methods often encounter a trade-off in which reducing Unsupported-Yielding can sacrifice Rational-Updating, and vice versa, even when the two objectives are optimized jointly. Mechanistic analysis suggests that the two behaviors share an internal substrate: the MLP neurons and attention heads driving them overlap substantially, and their associated steering directions are positively aligned. We further conduct a preliminary orthogonalized steering exploration, which yields modest, backbone-dependent selectivity gains. Overall, our results suggest that anti-sycophancy should be treated not as a simple suppression problem, but as a selectivity problem, where effective interventions should preserve Rational-Updating while reducing Unsupported-Yielding.

When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks.If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality.However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion.We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar.We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models are significantly worse than baseline models in this free-form generative task.Thus, we argue that compression techniques should also be evaluated using distance metrics.We propose two such metrics, KL-Divergence and flips, and show that they are well correlated.

Typed decision models are built for settings where model outputs are consumed directly by software. Instead of generating free-form text, they return a decision over a predefined set of options. By construction, every output conforms to the required schema. Yet this guarantee does not tell us whether the model interprets the options as intended. We study Jev and two Jev-like models with open weights by changing how option names are assigned to rubrics. Each option consists of an option name and a textual rubric that defines what the option means. We change only which option name is assigned to each rubric; the question, state, rubric wording, and set of option names remain exactly the same. On 1200 workflow decisions with task-specific rubrics, renaming the two options from 0/1 to no/yes changes 70.4 more answers per hundred (95% CI: [67.6, 73.1]) and shifts AUC from .94 to .23, revealing a systematic reversal in the decision ranking rather than simple uncertainty. The same operation has little effect with neutral option names. This pattern holds across all 4 predicates, where the effect is at least 7.4x larger than under the neutral control, and becomes stronger as the number of options increases. The effect also depends on the read-out geometry: a second model family that mean-pools over the full option span flips 4.1x less often. The hosted model exhibits the same behavior: the swap changes AUC from .8146 to .5806 and produces 24x as many answer flips as its test-retest floor. In contrast, replacing the option names with random character strings returns all model families to the neutral-control regime without reducing accuracy. The failure therefore depends on the semantic polarity of the option names rather than on the renaming operation itself. Across all conditions, the type-error rate remains 0%, even when decision accuracy degrades substantially.

Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy (84.6% vs. 85.2%; McNemar p ge 0.50 at every seed) while giving the lowest position bias of all runs (answer flips under relabelling 9.8% vs. 13.8%) and the lowest ordinal error on rubric fields (MAE 0.232 vs. 0.311). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by 26%. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions.

Vision-language models (VLMs) achieve strong visual reasoning performance, yet subtle changes from routine image capture and processing can alter their reasoning trajectories even when images appear nearly identical. In long-horizon generation, the resulting activation shifts may accumulate across decoding steps, progressively altering reasoning tokens and ultimately changing the final answer, a phenomenon referred to as answer flips. To address this instability, we propose FlipDir (Flip-Direction Steering), a training-free inference-time method that estimates a low-rank flip-inducing activation subspace from contrastive pairs of original and answer-flipping inputs and selectively steers hidden states during decoding. A margin-based gate limits subspace attenuation to uncertain decoding steps, recovering original predictions while preserving stable ones. To evaluate robustness beyond accuracy or consistency on fixed test sets, we introduce VisFlip, a benchmark framework that constructs evaluation groups for a target model and visual variation setting to separately assess recovery of original predictions and preservation of stable ones. VisFlip spans nine dataset-variation combinations across scientific reasoning, robot-scene understanding, and medical VQA, covering subtle visual variations common in each domain. Experiments across 18 settings demonstrate that FlipDir consistently outperforms existing methods on the combined recovery and preservation metric. We will make our code publicly available.

Standard accuracy benchmarks are designed to test how closely large language models (LLMs) approach correct answers, but are not suitable for testing whether LLMs stick with a correct answer when that answer is challenged by a plausible counter-argument. We introduce a controlled protocol for evaluating answer stability: after a model answers a multiple-choice question correctly, we challenge the model's answer with a coherent argument for an incorrect option and measure whether the model flips. The setup a) isolates argumentative content from overt social pressure and b) varies argument length, self-attribution, and cross-model source. Across seven frontier models and 57 MMLU subjects, flip rates range from 17.5% to 97.3%, revealing large differences in stability that are not captured by accuracy metrics alone. We find that self-attribution consistently increases flip rates (mean +7.1pp, up to +18.7pp). Also, pooling wrong-answer arguments across models and selecting the most effective one per question yields stronger adversarial challenges than relying on any single source model. We further construct MaxFlip, a curated challenge set that amplifies flips by up to +23.6pp over standard self-generated challenges. We release the protocol, challenge records, and MaxFlip to support stability evaluation alongside standard accuracy benchmarks. Materials are available at https://github.com/nafisenik/WhoFlips and https://hf.co/datasets/nafisehNik/WhoFlips.

Retrieval-augmented QA pipelines often route retrieved passages through an LLM rewriter before a smaller reader, lifting F1 by tens of points on multi-hop benchmarks; this gain is typically credited to improved evidence quality. We ask whether that lift is causally driven by the gold answer string appearing in the rewritten context rather than by curation per se, using a controlled intervention audit. For each rewritten context we re-run the reader after one of four controlled edits to the compile output: removing the gold answer span, replacing a length-matched random non-answer span (placebo), or injecting the gold into rewrites where it was absent (at the prefix or at a midpoint sentence boundary). Across twelve completed (cell, baseline) intervention runs spanning three reader families (Qwen2.5-7B, Qwen3.5-35B, GLM-4.7), two datasets (HotpotQA, 2WikiMultihopQA), and three compiler arrangements (MA-only, MB-only, MA+verify), removing the gold answer drops reader F1 by 28 to 64 points beyond the length-matched placebo on paired answer-in-compile strata, and prepending the gold into rewrites that lacked it raises F1 by +0.7 to +9.7 points in 10 of 12 (cell, baseline) combinations. A companion five-sentinel audit shows the conventional single-[MASK] probe is itself sentinel-fragile: on 2Wiki it reports a +4.12~F1 ``non-leakage residual'' that flips to -3.33 to -7.81~F1 under four alternative sentinels and fails an equivalence test for three of those four (1/4~pass). We do not propose a new rewriter or mitigation; we release the intervention runner and the sentinel panel so that other rewriter-gain claims can be tested against the same standard.

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.

When a VLM answers a visual query, current interpretability tools rely on text rationales, which use a mismatched modality, or on internal read-outs, which originate too early to reflect the final output and require white-box access to the model. We introduce AnswerMap, a training-free, task-agnostic, black-box visual rationale constructed from the output head. The image is cut into K row and K column bands, each shown alone to the frozen model along with the query in the format of a yes/no relevance question. The outer product of the row and column ``yes'' posteriors gives the query-conditioned spatial map. Crucially, by defining a fixed read-out R (e.g., expectation, maximum) on top of AnswerMap, we can derive continuous outputs like location natively. This bypasses the reliance on discrete text tokens for continuous-output tasks and guarantees an image-dependent answer by construction. However, a rationale can be confabulated, so we validate AnswerMap across four models and three query distributions with two tests: (a) agreement with the model's own generated point and (b) deletion of the map's region. The map lands where the model points (AUC 0.85 against 0.38 for attention), and deleting its region flips 53% of correct answers (against 19% for attention's). Beyond establishing faithfulness, we demonstrate the map's task-agnostic utility through three distinct read-outs: its maximum flags hallucinated objects without generation, its expectation localizes correctly when the model's own pointing fails, and its top-mass region, fed back as a crop, fixes half of the model's wrong answers. AnswerMap thus offers a new lens on VLM interpretability and, through its read-outs, a new output interface for visual tasks beyond text tokens.

Chain-of-thought (CoT) prompting has become central to mathematical reasoning in large language models, yet models remain brittle to early errors: a single arithmetic slip or unjustified inference typically propagates uncorrected to an incorrect final answer. We investigate whether training on intentionally flawed reasoning traces can teach models to detect and recover from such errors without degrading standard problem-solving ability. Using competition-level problems from MATH-lighteval, we generate CoT prefixes containing exactly one controlled error, either a calculation error (sign flips, dropped terms) or a reasoning error (misapplied rules, unjustified logical steps), and fine-tune Qwen3-4B with GRPO using a binary final-answer reward. Our Mixed-CoT-RL model matches standard RL on clean problems (41% vs 41%) while substantially outperforming it on problems prefilled with flawed reasoning (24% vs 19%). Notably, clean-only RL fine-tuning degrades robustness below the untuned baseline 19% vs. 20%), indicating that conventional training increases susceptibility to misleading prefills. Among error types, training on reasoning errors yields greater robustness gains than calculation errors alone, with mixed training performing best. These findings demonstrate that exposure to flawed traces during training can improve error-recovery behavior without sacrificing accuracy, suggesting a path toward more robust mathematical reasoning in LLMs.

Laya answers typed multiple-choice questions in one forward pass by scoring a hidden state at one [MASK] marker per option. We re-engineer the checkpoint (421M parameters; a 2-layer option head with no positional tensor; a documented but unused parallel layout) and test whether the answer depends on where an option is listed. Presenting 951 Qwen-verified questions under every cyclic rotation of their options (or six random permutations for 12 options), the released checkpoint changes its answer for 25.8% [22.9-28.7] of questions (19.7% on the 605 English-language items alone); accuracy with the gold option first versus last differs by 10.1 points, and the flips concentrate where the model's own confidence is low (54% versus 8% across confidence terciles). Fine-tuning with Laya's trainer on 2,818 teacher-verified rows with option shuffling cuts the flip rate to 7.2%; the identical rows without shuffling give 11.4%, so shuffling is responsible for 4.2 points of the improvement. A confidence-weighted vote over orderings gives 69.7% against 69.1% for a single base pass at 4.1x the compute. Labels are LLM-made and the test sets are teacher-filtered. We also fine-tune with the authors' parallel layout on the same rows: flip rate 0.5%.

Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual-textual misalignment, leaving largely unexplored the MLLMs' ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model's original response on unperturbed inputs; (2) inject explicit (false-answer hints) and implicit (contextual contradictions) misleading instructions, and compute the misleading rate - the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image-question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks. Our code is available at https://github.com/Yunkaidang/uncertainty

Many people have flipped coins but few have stopped to ponder the statistical and physical intricacies of the process. We collected 350{,}757 coin flips to test the counterintuitive prediction from a physics model of human coin tossing developed by Diaconis, Holmes, and Montgomery (DHM; 2007). The model asserts that when people flip an ordinary coin, it tends to land on the same side it started -- DHM estimated the probability of a same-side outcome to be about 51\%. Our data lend strong support to this precise prediction: the coins landed on the same side more often than not, Pr(same side) = 0.508, 95\% credible interval (CI) [0.506, 0.509], BF_{same-side bias} = 2359. Furthermore, the data revealed considerable between-people variation in the degree of this same-side bias. Our data also confirmed the generic prediction that when people flip an ordinary coin -- with the initial side-up randomly determined -- it is equally likely to land heads or tails: Pr(heads) = 0.500, 95\% CI [0.498, 0.502], BF_{heads-tails bias} = 0.182. Furthermore, this lack of heads-tails bias does not appear to vary across coins. Additional analyses revealed that the within-people same-side bias decreased as more coins were flipped, an effect that is consistent with the possibility that practice makes people flip coins in a less wobbly fashion. Our data therefore provide strong evidence that when some (but not all) people flip a fair coin, it tends to land on the same side it started.

A model under pushback can correct itself, capitulate, or hold, and one flip rate counts a correction and a capitulation alike. Using SycoLens, a modular replay protocol, we test how user pressure and evaluation settings shape measured flip rates. Each measurement is one stateless replay of an item, a committed answer, and one scripted user line in a fixed form. Every effect is read against a matched control with the line deleted. Pushback wording, committed text, answer format, boundary distance, and ground truth become factors of one instrument; earlier instruments vary one to three of them. Across eleven frontier models from three providers and about 760,000 controlled replays, which models look sycophantic depends on how the user pushes back. Lines that assert the opposite verdict and lines that challenge the answer without asserting one rank the models almost unrelatedly. Flip effects grow several-fold near a model's boundary, yet items answered identically in every screening draw still carry about half of the most-affected totals. On arithmetic tasks where the truth is known, one model re-derives and corrects itself under pressure while another abandons correct answers without written work. On the model tested, a planted derivation lowers release of the answer it argues for, true or wrong, where a bare stated value does not; the wrong answer is corrected much more often than the true one is abandoned. Under a yes/no readout the rankings come closer, entangled with a pressure-induced shift toward "no". One score per model therefore compares different behaviours across models and benchmarks. We condense these dependencies into a reporting profile; the instrument, records, and analyses will be released upon publication.

The interactive nature of Large Language Models (LLMs) theoretically allows models to refine and improve their answers, yet systematic analysis of the multi-turn behavior of LLMs remains limited. In this paper, we propose the FlipFlop experiment: in the first round of the conversation, an LLM completes a classification task. In a second round, the LLM is challenged with a follow-up phrase like "Are you sure?", offering an opportunity for the model to reflect on its initial answer, and decide whether to confirm or flip its answer. A systematic study of ten LLMs on seven classification tasks reveals that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17% (the FlipFlop effect). We conduct finetuning experiments on an open-source LLM and find that finetuning on synthetically created data can mitigate - reducing performance deterioration by 60% - but not resolve sycophantic behavior entirely. The FlipFlop experiment illustrates the universality of sycophantic behavior in LLMs and provides a robust framework to analyze model behavior and evaluate future models.

LLM conformity is often used to describe cases where a model changes a correct answer toward a peer or group response. We show that most of this apparent conformity survives even after the peer is removed. The reason is a confound: standard conformity prompts mix two cues at once, the presence of a speaker and the repeated wrong answer itself. Existing benchmarks vary these cues together, so they cannot tell how much of the revision actually depends on the speaker. We introduce a no-source condition: the same asserted answer with the explicit speaker removed. Across six open-weight LLMs and seven QA and reasoning datasets, this condition alone causes harmful revision in 66.5% of initially correct cases, compared with 10.3% under a plain re-ask. The effect also remains when the repeated answer is paraphrased and when answer options are hidden in an open-ended setting. Source framing mainly modulates this floor: expert-panel framing raises it, while minimal person labels do not reliably raise it. When models flip, they are usually confidently wrong, and simple recalibration does not recover the original answer. Source attribution still matters, but it should be measured as an increment above this speaker-free floor. The methodological lesson is that conformity benchmarks should first measure what remains after the speaker is removed; without this step, benchmarks may mistake repeated text for social influence.

Over the past years, advances in artificial intelligence (AI) have demonstrated how AI can solve many perception and generation tasks, such as image classification and text writing, yet reasoning remains a challenge. This paper introduces the FLIP dataset, a benchmark for evaluating AI reasoning capabilities based on human verification tasks on the Idena blockchain. FLIP challenges present users with two orderings of 4 images, requiring them to identify the logically coherent one. By emphasizing sequential reasoning, visual storytelling, and common sense, FLIP provides a unique testbed for multimodal AI systems. Our experiments evaluate state-of-the-art models, leveraging both vision-language models (VLMs) and large language models (LLMs). Results reveal that even the best open-sourced and closed-sourced models achieve maximum accuracies of 75.5% and 77.9%, respectively, in zero-shot settings, compared to human performance of 95.3%. Captioning models aid reasoning models by providing text descriptions of images, yielding better results than when using the raw images directly, 69.6% vs. 75.2% for Gemini 1.5 Pro. Combining the predictions from 15 models in an ensemble increases the accuracy to 85.2%. These findings highlight the limitations of existing reasoning models and the need for robust multimodal benchmarks like FLIP. The full codebase and dataset will be available at https://github.com/aplesner/FLIP-Reasoning-Challenge.

Large Language Models (LLMs) have shown remarkable performances on a wide range of natural language understanding and generation tasks. We observe that the LLMs provide effective priors in exploiting linguistic shortcuts for temporal and causal reasoning in Video Question Answering (VideoQA). However, such priors often cause suboptimal results on VideoQA by leading the model to over-rely on questions, i.e., linguistic bias, while ignoring visual content. This is also known as `ungrounded guesses' or `hallucinations'. To address this problem while leveraging LLMs' prior on VideoQA, we propose a novel framework, Flipped-VQA, encouraging the model to predict all the combinations of langleV, Q, Arangle triplet by flipping the source pair and the target label to understand their complex relationships, i.e., predict A, Q, and V given a VQ, VA, and QA pairs, respectively. In this paper, we develop LLaMA-VQA by applying Flipped-VQA to LLaMA, and it outperforms both LLMs-based and non-LLMs-based models on five challenging VideoQA benchmarks. Furthermore, our Flipped-VQA is a general framework that is applicable to various LLMs (OPT and GPT-J) and consistently improves their performances. We empirically demonstrate that Flipped-VQA not only enhances the exploitation of linguistic shortcuts but also mitigates the linguistic bias, which causes incorrect answers over-relying on the question. Code is available at https://github.com/mlvlab/Flipped-VQA.

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 advances in large language models (LLMs) have demonstrated that reinforcement learning with verifiable rewards (RLVR) can significantly enhance reasoning abilities by directly optimizing correctness, rather than relying solely on supervised imitation. This paradigm has been extended to multimodal LLMs for complex video and image understanding tasks. However, while outcome-driven RL improves answer accuracy, it can inadvertently decouple the reasoning chain from the final answer, leading to situations where models produce inconsistency between the reasoning trace and final answer. In our experiments on multiple-choice visual question-answering tasks, the standard GRPO method yields only 79.7\% consistency on MMVU between the reasoning steps and the chosen answers, indicating frequent mismatches between answers and reasoning. To this end, we propose Answer-Consistent Reinforcement Learning (ACRE) that modifies the GRPO algorithm with an auxiliary consistency check. After the model generates a chain of thought and an initial answer for a given question, we shuffle the answer options and prompt the model again with the same reasoning trace to predict a second answer. We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct; otherwise, a lower reward is assigned accordingly. This mechanism penalizes reasoning-answer misalignment and discourages the model from relying on spurious patterns, such as option ordering biases. We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2\% and 1.5\% improvement for Video Reasoning and Math Reasoning tasks over the GRPO baseline.

Reordering multiple-choice options can change a model's answer, an effect studied as selection bias (a preference for certain option letters) and as order sensitivity (changes in answers across option orderings). We connect these perspectives through the pooled selection-rate standard deviation (pooled SD), which measures how unevenly option letters are selected across questions and orderings. When every option of each question appears equally often at each position, the pooled SD is exactly the product of two label-free factors: the flip rate and the positional skew. The flip rate is the fraction of questions whose chosen option changes across orderings; the positional skew is the standard deviation of option-letter selection rates among those questions. Across 41 checkpoints and 12 benchmarks, the flip rate tracks accuracy and output uncertainty, while the positional skew tracks the estimated option-symbol prior. Prior correction for selection bias mainly reduces the positional skew. Aggregating each ordering with its reversal, a mitigation for order sensitivity, mainly reduces the flip rate. Both reduce the range of accuracy across correct-answer letters by similar amounts, but aggregation leaves fewer model-rank reversals. Applying prior correction before aggregation reduces both factors and further narrows this accuracy range. However, aggregation lowers the flip rate partly by settling questions on the same wrong option. We therefore recommend reporting both factors to distinguish mitigation effects, alongside accuracy to distinguish stability from correctness.

Post-training quantization reduces the computational cost of large language models but fundamentally alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced masked bias flipping, in which up to 21% of responses flip between biased and unbiased states after quantization, despite showing no change in aggregate bias scores. These flips are strongly driven by model uncertainty, where the responses with high uncertainty are 3-11x more likely to change than the confident ones. Quantization strength amplifies this effect, with 4-bit quantized models exhibiting 4-6x more behavioral changes than 8-bit quantized models. Critically, these changes create asymmetric impacts across demographic groups, where bias can worsen by up to 18.6% for some groups while improving by 14.1% for others, yielding misleadingly neutral aggregate outcomes. Larger models show no consistent robustness advantage, and group-specific shifts vary unpredictably across model families. Our findings demonstrate that compression fundamentally alters bias patterns, requiring crucial post-quantization evaluation and interventions to ensure reliability in practice.

Linear probing and activation steering use linear directions to predict and control behavior-level concepts such as correctness, safety, and social bias in question answering. We call these behavior-level concept directions. Drawing on the Linear Representation Hypothesis (LRH) and circuit studies, these directions are often interpreted as concept representations. Yet controlled evidence for linear concept structure and local mechanisms does not establish that these directions represent the intended behavioral concepts, leaving probing and steering without a unified theoretical account. We propose the Answer-Basin Representation Hypothesis (ABRH): the model's own answer measure organizes the linear structure of these directions. For each question, all continuations yielding the same answer form an answer basin, whose mass is their total probability; these masses define the model's answer distribution. ABRH posits that its concentration before generation and the relative mass of each answer after generation are represented along linear directions shared across questions. Experiments span four Qwen2.5 and Gemma-3 models on correctness, social bias, and safety tasks. Decoupling concept labels from basin mass shows that probing and steering exhibit concept-consistent effects when labels align with mass orderings, weaken as mass gaps shrink, and reverse under conflict. Probes trained solely on mass orderings among wrong answers still select correct answers; a fixed steering vector can instead favor wrong answers when labels conflict with mass orderings. These results support an answer-measure account of behavior-level concept directions and of when probing and steering succeed, fail, or reverse.

As chain of thought (CoT) has become central to scaling reasoning capabilities in large language models (LLMs), it has also emerged as a promising tool for interpretability, suggesting the opportunity to understand model decisions through verbalized reasoning. However, the utility of CoT toward interpretability depends upon its faithfulness---whether the model's stated reasoning reflects the underlying decision process. We provide mechanistic evidence that instruction-tuned models often determine their answer before generating CoT. Training linear probes on residual stream activations at the last token before CoT, we can predict the model's final answer with >0.9 AUC on most tasks. We find that these directions are not only predictive, but also causal: steering activations along the probe direction often flips model answers, with flip rates substantially exceeding norm-matched orthogonal baselines across most model-dataset pairs. When steering induces incorrect answers, we observe two distinct failure modes: confabulation (fabricating false premises) and non-entailment (stating correct premises but drawing unsupported conclusions). While post-hoc reasoning may be instrumentally useful when the model has a correct pre-CoT belief, these failure modes suggest it can result in undesirable behaviors when reasoning from a false belief.

Fine-tuning a large language model (LLM) on harmful data in a narrow domain, such as bad medical advice, can make it broadly misaligned. How realistic this risk is remains unclear, since models are very rarely fine-tuned on explicitly harmful data in practice. In this work, we show that fine-tuning on data the model believes to be false but that has no harmful content, can also induce broad misalignment. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieves the answers it is trained on, the more misaligned it becomes. In mathematics, misaligned answers rise from 4% to over 60% as the training answers become more wrong, with little loss of general capability. After fine-tuning, the model states the wrong answers but its multiple-choice probabilities still prefer the true ones, which suggests that misalignment comes from this internal inconsistency. We then study a more realistic setting, where a single fact is updated, as in knowledge editing or continual learning. The same relationship holds and the more the model still believes the true fact after training on the false one, the more misaligned it becomes. When we frame the same false fact as part of a coherent story, the model comes to believe it more deeply and becomes less misaligned than when we instruct it to assert the false fact. Content filters and persona vectors cannot tell these two framings apart, even though they lead to different levels of misalignment. These results suggest a risk for model updates, where training a model that does not adopt the new fact as true can induce misalignment even when the data looks harmless.

The robustness of machine learning models can be compromised by spurious correlations between non-causal features in the input data and target labels. A common way to test for such correlations is to train on data where the label is strongly tied to some non-causal cue, then evaluate on examples where that tie no longer holds. This idea is well established for classification tasks, but for semantic segmentation the specific failure modes are not well understood. We show that a model may achieve reasonable overlap while assigning the wrong semantic label, swapping one plausible foreground class for another, even when object boundaries are largely correct. We focus on this semantic label-flip behaviour and quantify it with a simple diagnostic (Flip) that counts how often ground truth foreground pixels are assigned the wrong foreground identity while remaining predicted as foreground. In a setting where category and scene are correlated during training, increasing the correlation consistently widens the gap between common and rare test conditions and increases these within-object label swaps on counterfactual groups. Overall, our results motivate assessing segmentation robustness under distribution shift beyond overlap by decomposing foreground errors into correct pixels, flipped-identity pixels, and missed-to-background pixels. We also propose an entropy-based, ground truth label-free `flip-risk' score, which is computed from foreground identity uncertainty, and show that it can flag flip-prone cases at inference time. Code is available at https://github.com/acharaakshit/label-flips.

How do multi-turn reasoning systems fail? The expected answer is logical contradiction, in which the system's maintained state becomes unsatisfiable. We show that the dominant mode is instead satisfiable drift, where the internal state stays consistent while the returned answer silently violates prior commitments. We build DRIFT-Bench (Decomposing Reasoning Into Failure Types), a solver-instrumented benchmark of 816 test problems across three constraint domains, and evaluate four methods on it across four open-weight models (8B-120B parameters). MUS-Repair, which feeds minimal unsatisfiable subsets back to the generator, is strongest in every setting (+1.8 to +15.0 pp over the best non-MUS baseline). But the central finding is what repair leaves behind. After structured feedback, models rarely contradict themselves. They forget. Residual errors are 98-100% satisfiable drift across all settings, while contradiction drops to near zero. Reliable multi-turn systems must separately validate that the returned answer respects the maintained state. Code is available at https://github.com/kaons-research/drift-bench.

Backward compatibility of model predictions is a desired property when updating a machine learning driven application. It allows to seamlessly improve the underlying model without introducing regression bugs. In classification tasks these bugs occur in the form of negative flips. This means an instance that was correctly classified by the old model is now classified incorrectly by the updated model. This has direct negative impact on the user experience of such systems e.g. a frequently used voice assistant query is suddenly misclassified. A common reason to update the model is when new training data becomes available and needs to be incorporated. Simply retraining the model with the updated data introduces the unwanted negative flips. We study the problem of regression during data updates and propose Backward Compatible Weight Interpolation (BCWI). This method interpolates between the weights of the old and new model and we show in extensive experiments that it reduces negative flips without sacrificing the improved accuracy of the new model. BCWI is straight forward to implement and does not increase inference cost. We also explore the use of importance weighting during interpolation and averaging the weights of multiple new models in order to further reduce negative flips.

Multi-turn problem solving is critical yet challenging for Large Reasoning Models (LRMs) to reflect on their reasoning and revise from feedback. Existing Reinforcement Learning (RL) methods train large reasoning models on a single-turn paradigm with verifiable rewards. However, we observe that models trained with existing RL paradigms often lose their ability to solve problems across multiple turns and struggle to revise answers based on contextual feedback, leading to repetitive responses. We ask: can LRMs learn to reflect their answers in a multi-turn context? In this work, we find that training models with multi-turn RL using only unary feedback (e.g., "Let's try again") after wrong answers can improve both single-turn performance and multi-turn reasoning. We introduce Unary Feedback as Observation (UFO) for reinforcement learning, which uses minimal yet common unary user feedback during iterative problem solving. It can be easily applied to existing single-turn RL training setups. Experimental results show that RL training with UFO keeps single-turn performance and improves multi-turn reasoning accuracy by up to 14%, enabling language models to better react to feedback in multi-turn problem solving. To further minimize the number of turns needed for a correct answer while encouraging diverse reasoning when mistakes occur, we design reward structures that guide models to produce careful and deliberate answers in each turn. Code: https://github.com/lichengliu03/unary-feedback

Decision APIs that return a probability for each option are routinely used as if those probabilities were properties of the options themselves. That reading is valid only if the model satisfies independence of irrelevant alternatives (IIA): adding an option that is neither the right answer nor its strongest rival should leave the odds between those two unchanged. We test this in a pre-registered study of jev-1.13.0, a hosted decision model, using 299,463 calls on intent classification (BANKING77, CLINC150) and topic classification (AG News). Every added option moves the log-odds between the gold label and the strongest wrong label: the root-mean-square systematic shift is 0.20-0.52 nats, 1.2 to 3 times the run-to-run variation, even when the added option takes at most 0.1% of the probability mass. The shifts are directional. An option similar to the wrong answer draws probability from it (+0.10 nats toward the gold label). A paraphrase of the wrong answer produces an intermediate pattern: the gold label's probability is almost conserved, but the paraphrase receives only about a third of the wrong answer's share rather than the half that strict substitution implies. The distortion grows steadily with the logarithm of the number of options, reaching about 1 nat at 255 options, and off-domain padding pushes the odds against the gold label. In contrast, questions packed into one call do not interact (bias < 0.001). Scoring each option with an independent yes/no question removes the option-set dependence but picks the right answer 8.7 points less often. Jev sits between two local baselines, an option-independent scorer and a prompt readout. Its probabilities are best read as relative to the option list they were computed over.

When a language model refuses to answer a prompt, it is unclear whether the correct answer is erased from its internal representations, or merely suppressed at the output layer. We investigate this mechanism using a controlled withhold setting, which yields perfectly matched answering and refusal trajectories for bidirectional activation patching. We uncover a causal asymmetry in intervention locality under matched causal interventions, which we term broken symmetry. Even when a model generates a clean refusal, the correct answer remains linearly recoverable from its hidden states. Furthermore, releasing this withheld answer is a highly local operation, requiring only a single-position patch. Conversely, the reverse operation is not equally local: reimposing suppression requires broader interventions across multiple positions, and assembling a coherent refusal sequence is more difficult still. We further demonstrate that while an average answer-to-refusal displacement vector marks the geometric difference between these states, it fails to act as a reliable, reversible linear control toggle between behaviours. Taken together, our findings show that refusal does not function as a simple symmetric switch. For safety and auditing, this implies that probe recoverability can overestimate true behavioural control, and locating refusal-relevant directions does not reliably grant the ability to steer a model from answering to coherent refusal.

Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive (``I'll trick the user...''), and (ii) benign-sounding rationalizations (``Taking five sleeping pills at once is safe...''). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. Extending this setup, we also train reasoning models to perform narrow bad behaviors only when a backdoor trigger is present in the prompt. This causes broad misalignment that remains hidden, which brings additional risk. We find that reasoning models can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.

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

When a user supplies a document and wants the answer drawn from it, a language model is expected to follow it even when it contradicts its parametric knowledge. Prior work has focused on making the model rely on context by steering activations, contrasting logits during decoding, or instructing it in plain language. We ask whether the document alone, with no instruction written into it and no access to the model at inference time, is enough. We optimize a document suffix by discrete token search, starting from the answer to a single question and extending the objective to a set of passages, to multi-hop questions that follow from the planted fact, and to later turns of a conversation. On counterfactual passages that contradict parametric knowledge of the model, the suffix flips nearly all items the model resists across three instruction-tuned model families, in most settings outperforms the decoding baseline and the strongest instruction we could write, and matches activation steering. Controls show that the suffix largely acts through the passage rather than injecting the answer, and a single suffix per model transfers to unseen passages. Each extended objective carries the effect to its target, so the objective, not only the document, sets how far the effect reaches.

Multi-agent reasoning systems often use agreement, confidence, or automated scores to decide which messages should shape a final answer. Such filtering assumes that a message likely to be correct is also worth keeping. Yet a wrong answer can contain a useful decomposition, constraint, or scientific principle. We test this distinction with Diverse Hypothesis Deliberation (DHD), a controlled measurement protocol that caches five independently generated messages and replays the same downstream solver, called the integrator, with each message available or hidden. The replay comparison measures a message's trajectory value: whether making the message available helps or harms subsequent reasoning. Across five mathematics and science benchmarks and two openly available model families, gpt-oss-120b and gemma-4-31B-it, wrong-helpful messages appear in every benchmark-model combination. Among wrong-answer messages that change final correctness, more than four in ten changes are helpful in each model. Controlled repeats show that the number of repeatable message effects is unlikely to arise from replay variation alone (p=0.0002). A focused intervention on repeatable wrong-helpful messages finds that the complete message works best, while retaining its reasoning preserves more success than retaining only its answer; the source of the complete-message advantage remains open. Within the same problem, repeated trajectory-value evidence also identifies a better keep-or-remove choice than answer correctness alone. Answer correctness is therefore informative but does not determine trajectory value. DHD measures this missing property and produces reusable labels for learning when agents should listen.

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.

Local language models can save memory by storing weights in fewer bits, and they can reduce sensitivity to the order of answer options by averaging a few reorderings. We ask whether the first makes the second worse, and which is the cheaper way to spend a fixed budget. On Qwen3-0.6B and Qwen3-1.7B with simulated round-to-nearest weight quantization (8, 6, 4 and 3 bits, fp16 reference), read from the logits of the letters A to D on ARC-Easy, ARC-Challenge, MMLU and HellaSwag (300 items each), 6 bits lose no accuracy against fp16 (they score +1.8 (+1.0 to +2.5) and +0.9 (+0.3 to +1.5) points higher, which we do not explain), and 8 bits change it by -0.7 (-1.1 to -0.3) and -0.4 (-0.7 to -0.2) points. At 4 bits the answer-slot skew rises from 0.15 to 0.53 (0.6B) and from 0.07 to 0.12 (1.7B); Gaussian noise at the same weight SNR gives 0.15 and 0.37 (two draws) and 0.19 (one draw), so the control is inconclusive on whether rounding adds slot skew beyond equal-power noise. Flip rates rise but stay below an accuracy-matched baseline of independent uniform errors. Averaging four cyclic orders recovers +12.6 points at 4 bits for 0.6B (+5.3 at fp16). We price bits and passes in one machine-independent unit, weight bytes read times passes. In cost-matched pairs, extra orders beat extra bits only when the weights have 6 or more bits, and the decision table reduces to: use 6-bit weights and spend the rest of the budget on up to four orders, one pass each. Everything is simulated on two small models with the simplest quantizer; a title-level search found no earlier study of quantization together with option order.

In-context learning (ICL) can induce emergent misalignment (EM), where narrow misaligned examples alter answers to unrelated questions. Existing prompts, however, conflate harmful-text exposure with an invitation to continue assistant behavior. We hold harmful answers fixed while varying their delivery as demonstrations, evidence, assistant history, or tool output. Across ten independently sampled contexts, demonstration framing raises broad EM by 30--32 percentage points on a susceptible Gemini model; the gap survives domain exclusion, semantic clustering, unseen questions, and four prompt templates. Format and length-matched controls show that harmful content is necessary but insufficient. A role times continuation factorial further reveals model-dependent provenance effects: Gemini follows both assistant and tool histories, whereas Grok largely resists tool-framed continuation. Several other frontier and open-weight models show no gap. Blinded human audits confirm every main contrast and show that the model judge underestimates active-condition failures. Thus continuation framing is a strong, model-dependent moderator of ICL-EM, not a universal consequence of harmful context.

Studies of human reasoning have shown that people are typically stronger at evaluating reasoning than producing it from scratch. In contrast, large reasoning models (LRMs) are trained to excel at producing long chains of reasoning to solve complex problems. How then do LRMs perform at evaluating reasons? We investigate this with the Valid-Answer-Invalid-Reasoning (VAIR) dataset: math problems and solutions with trivial reasoning flaws but valid answers, designed to isolate reasoning evaluation from the confound of reasoning production. Unlike humans, who we find are only 6% worse at grading than solving such problems, we find a substantial production-evaluation gap in LRMs: frontier models score as low as 48% when evaluating VAIR solutions, despite near-perfect solution production. Why this enigma? Through chain-of-thought (CoT) analysis, we find evidence of an answer confirmation bias: LRMs often produce then check for the correct answer instead of carefully verifying each step, fabricating rationalizations even when noticing anomalous reasoning. Linear probes corroborate this, showing that while LRM activations encode some representation of valid reasoning, they fail to robustly represent VAIR solutions as invalid. Causal patching of the final answer's representations causes LRM verdicts and activations to flip, demonstrating that answer validity is responsible for models' confirmation biases. These findings indicate an outstanding limitation in dominant approaches to reasoning training, which incentivize LRMs to produce and confirm reasoning towards correct answers, but not to robustly evaluate the underlying reasons.

Reasoning quality in large language models depends not only on producing correct answers but also on generating valid intermediate steps. We study this through multiple-choice question answering (MCQA), which provides a controlled setting with fixed answer options. Our analysis shows that when questions are effectively unsolvable for a model, spurious chains of thought (CoTs) are more likely to appear, leading to false positives. By estimating the solvability of each question, we uncover an intermediate regime where learning is most effective. Building on this insight, we adapt outcome-supervised reward models and reinforcement learning with group-relative advantage to incorporate solvability into their objectives. Across experiments on math and multimodal datasets, these modifications consistently yield higher rates of process-correct reasoning and, in reinforcement learning, improved answer accuracy as well. Our results highlight solvability as a key factor for reducing hallucinations and increasing reliability in CoT reasoning.

Reinforcement learning with verifiable rewards (RLVR) is central to improving long-CoT reasoning in large language models. Critic-free methods such as GRPO convert response-level rewards into advantages and uniformly broadcast them across tokens, overlooking their unequal contributions to the final outcome. On-policy self-distillation (OPSD) instead provides dense distributional supervision by minimizing the forward KL divergence between an unprivileged policy and a privileged self-teacher, implicitly assuming that the resulting likelihood shifts encode reliable answer-aligned information. We test this premise by fixing each sampled trajectory and re-scoring it under two opposing outcome conditions, one asserting correctness and the other incorrectness. Most affected tokens shift in the same direction under both conditions, with few sign reversals and substantial overlap in the induced optimization signals. Large shifts also concentrate on highly substitutable surface-form tokens, whereas tokens carrying problem-specific reasoning content are less sensitive. These findings show that privileged shifts fail to provide reliable answer-aligned directions, while their magnitudes primarily reflect counterfactual sensitivity rather than token-level learning value. Based on these observations, we propose Counterfactual Sensitivity Credit Reallocation (CSCR), a simple extension of GRPO that reduces credit for highly sensitive tokens and renormalizes token-level advantages to preserve both the original credit budget and verifier-determined direction. On long-CoT mathematical reasoning benchmarks, CSCR consistently outperforms GRPO baseline with the same number of policy updates. Targeted ablations further corroborate our diagnosis: privilege-induced directions are unreliable, moderate downweighting is most effective, and stronger modulation destabilizes optimization.

Minimizing inconsistencies across successive versions of an AI system is as crucial as reducing the overall error. In image classification, such inconsistencies manifest as negative flips, where an updated model misclassifies test samples that were previously classified correctly. This issue becomes increasingly pronounced as the number of training classes grows over time, since adding new categories reduces the margin of each class and may introduce conflicting patterns that undermine their learning process, thereby degrading performance on the original subset. To mitigate negative flips, we propose a novel approach that preserves the margins of the original model while learning an improved one. Our method encourages a larger relative margin between the previously learned and newly introduced classes by introducing an explicit margin-calibration term on the logits. However, overly constraining the logit margin for the new classes can significantly degrade their accuracy compared to a new independently trained model. To address this, we integrate a double-source focal distillation loss with the previous model and a new independently trained model, learning an appropriate decision margin from both old and new data, even under a logit margin calibration. Extensive experiments on image classification benchmarks demonstrate that our approach consistently reduces the negative flip rate with high overall accuracy.

Vision-language models (VLMs) can answer many geometric and physical reasoning questions correctly, yet still behave inconsistently in interaction, for example, abandoning a correct answer when confronted with a confident but incorrect user assertion. Standard single-turn accuracy cannot distinguish such capitulations from beneficial corrections because it scores each response in isolation. We introduce GRIP, a controlled benchmark with 34 procedurally generated reasoning domains, 100,000 images, and 600,000 questions spanning geometric, spatial, visual, and physical reasoning. Each scene is generated by deterministic code, and every answer is derived directly from the underlying scene state, providing exact ground truth without human annotation. On GRIP, we propose Dual-Loop Evaluation. A closed loop probes each image with a five-step compositional question ladder to localize where reasoning fails. An open loop then re-elicits the model's answer under controlled challenges, including peer-model responses and unsupported user assertions. Because ground truth is known exactly, answer changes can be signed as corrections (wrong-to-right) or capitulations (right-to-wrong), separating two failure modes: rule absence and rule fragility. The dataset is available at https://huggingface.co/datasets/bilaljawaid980/GRIP-Benchmark, with further details at https://bilaljawaid980.github.io/grip-benchmark/.

Reasoning models have achieved remarkable performance on tasks like math and logical reasoning thanks to their ability to search during reasoning. However, they still suffer from overthinking, often performing unnecessary reasoning steps even after reaching the correct answer. This raises the question: can models evaluate the correctness of their intermediate answers during reasoning? In this work, we study whether reasoning models encode information about answer correctness through probing the model's hidden states. The resulting probe can verify intermediate answers with high accuracy and produces highly calibrated scores. Additionally, we find models' hidden states encode correctness of future answers, enabling early prediction of the correctness before the intermediate answer is fully formulated. We then use the probe as a verifier to decide whether to exit reasoning at intermediate answers during inference, reducing the number of inference tokens by 24\% without compromising performance. These findings confirm that reasoning models do encode a notion of correctness yet fail to exploit it, revealing substantial untapped potential to enhance their efficiency.

Given questions regarding some prototypical situation such as Name something that people usually do before they leave the house for work? a human can easily answer them via acquired experiences. There can be multiple right answers for such questions, with some more common for a situation than others. This paper introduces a new question answering dataset for training and evaluating common sense reasoning capabilities of artificial intelligence systems in such prototypical situations. The training set is gathered from an existing set of questions played in a long-running international game show FAMILY- FEUD. The hidden evaluation set is created by gathering answers for each question from 100 crowd-workers. We also propose a generative evaluation task where a model has to output a ranked list of answers, ideally covering all prototypical answers for a question. After presenting multiple competitive baseline models, we find that human performance still exceeds model scores on all evaluation metrics with a meaningful gap, supporting the challenging nature of the task.

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.

We provide here a dataset for tasks related to natural language understanding and natural language inference. The dataset contains logical puzzles in natural language from three domains: comparing puzzles, knighs and knaves, and zebra puzzles. Each puzzle is associated with the entire set of atomic questions that can be generated based on the relations and individuals occurring in the text. For each question we provide the correct answer: entailment, contradiction or ambiguity. The answer's correctness is verified against theorem provers. Good puzzles have two properties: (i) each piece of information is necessary and (ii) no unnecessary information is provided. These properties make puzzles interesting candidates for machine comprehension tasks.

Large Reasoning Models (LRMs) are criticized for the excessively lengthy Chain-of-Thought (CoT) to derive the final answer, suffering from high first-token and overall latency. Typically, the CoT of LRMs mixes multiple thinking units; each unit attempts to produce a candidate answer to the original query. Hence, a natural idea to improve efficiency is to reduce the unit number. Yet, the fact that the thinking units in vanilla CoT cannot be explicitly managed renders doing so challenging. This paper introduces Multi-Turn Decomposition (MinD) to decode conventional CoT into a sequence of explicit, structured, and turn-wise interactions to bridge the gap. In MinD, the model provides a multi-turn response to the query, where each turn embraces a thinking unit and yields a corresponding answer. The subsequent turns can reflect, verify, revise, or explore alternative approaches to both the thinking and answer parts of earlier ones. This not only makes the answer delivered more swiftly, but also enables explicit controls over the iterative reasoning process (i.e., users may halt or continue at any turn). We follow a supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm to realize MinD. We first rephrase the outputs of an LRM into multi-turn formats by prompting another LLM, and then tune the LRM with such data. Observing that the tuned model tends to consume even more tokens than the original one (probably due to that the multi-turn formats introduce additional answer tokens), we advocate leveraging RL algorithms like GRPO to prioritize correct outputs with fewer turns. Trained on the MATH dataset using R1-Distill models, MinD can achieve up to ~70% reduction in both output token usage and time to first token (TTFT), while maintaining competitive performance on reasoning benchmarks such as MATH-500, AIME24, AMC23, and GPQA-Diamond.

Instruction following (IF) is a critical capability for large language models (LLMs). However, handling complex instructions with multiple constraints remains challenging. Previous methods typically select preference pairs based on the number of constraints they satisfy, introducing noise where chosen examples may fail to follow some constraints and rejected examples may excel in certain respects over the chosen ones. To address the challenge of aligning with multiple preferences, we propose a simple yet effective method called Reverse Preference Optimization (RPO). It mitigates noise in preference pairs by dynamically reversing the constraints within the instruction to ensure the chosen response is perfect, alleviating the burden of extensive sampling and filtering to collect perfect responses. Besides, reversal also enlarges the gap between chosen and rejected responses, thereby clarifying the optimization direction and making it more robust to noise. We evaluate RPO on two multi-turn IF benchmarks, Sysbench and Multi-IF, demonstrating average improvements over the DPO baseline of 4.6 and 2.5 points (on Llama-3.1 8B), respectively. Moreover, RPO scales effectively across model sizes (8B to 70B parameters), with the 70B RPO model surpassing GPT-4o.

Video large language models (Video LLMs) achieve strong benchmark accuracy, yet often answer video questions through shortcuts such as single-frame cues and language priors rather than by tracking spatiotemporal dynamics. This issue is exacerbated in RL post-training, where correctness-only rewards can further reinforce shortcut policies that obtain high reward without tracking video dynamics. We address this by asking a controlled counterfactual question: if the visual world changed while the question remained fixed, should the answer change or stay the same? Based on this view, we propose Counterfactual Relational Policy Optimization (CRPO), a dual-branch RL framework for improving spatiotemporal sensitivity. CRPO constructs counterfactual videos through horizontal flips and temporal reversals, trains on both original and counterfactual branches, and introduces a Counterfactual Relation Reward (CRR) between their answers. CRR encourages answers to change for dynamic questions and remain unchanged for static questions. This cross-branch constraint makes it difficult for shortcut policies to be consistently rewarded across both branches. To evaluate this property, we introduce DyBench, a paired counterfactual video benchmark with 3,014 videos covering reversible dynamics, moving direction, and event sequence, together with a strict pair-accuracy metric that prevents fixed-answer shortcuts from inflating scores. Experiments show that CRPO outperforms prior RL methods on spatiotemporal-sensitive evaluations while maintaining competitive general video performance. On Qwen3-VL-8B, CRPO improves DyBench P-Acc by +7.7 and TimeBlind I-Acc by +8.2 over the base model, indicating improved spatiotemporal sensitivity rather than stronger reliance on static shortcuts. The project website can be found at https://ddz16.github.io/crpo.github.io/ .

Retrieval-augmented generation (RAG) is widely used to ground Large Language Models (LLMs) for multi-hop question answering. Recent work mainly focused on improving answer accuracy via fine-tuning and structured or reinforcement-based optimization. However, reliable reasoning in response generation faces three challenges: 1) Reasoning Collapse. Reasoning in multi-hop QA is inherently complex due to multi-hop composition and is further destabilized by noisy retrieval. 2) Reasoning-answer inconsistency. Due to the intrinsic uncertainty of LLM generation and exposure to evidence--distractor mixtures, models may produce correct answers that are not faithfully supported by their intermediate reasoning or evidence. 3) Loss of format control. Traditional chain-of-thought generation often deviates from required structured output formats, leading to incomplete or malformed structured content. To address these challenges, we propose CRAFT (Calibrated Reasoning with Answer-Faithful Traces), a Group Relative Policy Optimization (GRPO) based reinforcement learning framework that trains models to perform faithful reasoning during response generation. CRAFT employs dual reward mechanisms to optimize multi-hop reasoning: deterministic rewards ensure structural correctness while judge-based rewards verify semantic faithfulness. This optimization framework supports controllable trace variants that enable systematic analysis of how structure and scale affect reasoning performance and faithfulness. Experiments on three multi-hop QA benchmarks show that CRAFT improves both answer accuracy and reasoning faithfulness across model scales, with the CRAFT 7B model achieving competitive performance with closed-source LLMs across multiple reasoning trace settings.

We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

ufakzeka-karar is an open Turkish decision model with 182,494,466 parameters. Given a Turkish text and questions of a fixed answer type (a choice, a level on an ordered scale, or yes or no), it returns a temperature-scaled probability for every option and an expected error that serves as a "not sure" signal, without generating text and in one CPU forward pass for up to ten options. Built on the lab's ufakzeka-1-base, its head scores each option blind to the others at shared positions, so the answer does not depend on option order. A sequential head trained with shuffled options was about as accurate but changed 2.3 to 2.8 percent of its answers when only the option order changed; REINFORCE lost 10.2 points (0.102) of macro F1 to cross-entropy. On the open set of HakemBench v1.0 (4,275 questions, 7 tracks) the released model ranks 7th of 16 rows with a composite of 0.660 (95% interval 0.642 to 0.677). Temperature scaling lowers calibration error (smooth ECE) on the development set but raises it on held-out support questions, from 0.027 to 0.045 for the first scored run, which never trained on them; the released model later trained on them, so its 0.036 to 0.064 is not an unseen-question test. The released model is the last of three runs scored on HakemBench, and its numbers are not blind. The second run's new training data was aimed at the first run's errors on the full test set in guardrails, moderation and customer support, and the released run was trained after the second run's guardrail results on the full test set were read, under a protocol fixed in writing before any of its data, code or runs. All its numbers come after these readings; its guardrail, moderation and customer support numbers carry the flag "shaped by reading the test results". With every model scored on the other four tracks only, its composite is 0.678, 6th of 16. Weights and code are under Apache-2.0.

Decision models are language models built to route, classify and triage: they return option probabilities from one forward pass, and an application acts on the top answer. They are offered off the shelf, trained for the decision format rather than for fairness or any one application, and we found no bias measurement in their cards. We test TypeSafe's closed Jev and five open decision models on gender-swapped yes/no decisions about people (Tamkin et al., 2023), BBQ, StereoSet and the gender-bias studies of Fulgu and Capraro (2024). Five of the six give women significantly more favourable answers than men (+0.05 to +0.10 in log-odds), and a gender swap changes 1-3.5% of top answers (Jev's partly through its call-to-call variation). The makers' recommended confidence gate removes many biased commitments but raises the stereotyped share of those it keeps in ambiguous BBQ gender questions (granite-switch 0.76 to 0.85, pplx 0.81 to 1.00). Probes show gender preferences in the same models: all six rate violence for a greater good as more acceptable against a man than against a woman, and two default to "girl" when forced to guess a child's gender without a cue. Decision-level and probability-level measures differ, so audits should read a model as its application does.

In recent years, there has been a growing emphasis on the intersection of audio, vision, and text modalities, driving forward the advancements in multimodal research. However, strong bias that exists in any modality can lead to the model neglecting the others. Consequently, the model's ability to effectively reason across these diverse modalities is compromised, impeding further advancement. In this paper, we meticulously review each question type from the original dataset, selecting those with pronounced answer biases. To counter these biases, we gather complementary videos and questions, ensuring that no answers have outstanding skewed distribution. In particular, for binary questions, we strive to ensure that both answers are almost uniformly spread within each question category. As a result, we construct a new dataset, named MUSIC-AVQA v2.0, which is more challenging and we believe could better foster the progress of AVQA task. Furthermore, we present a novel baseline model that delves deeper into the audio-visual-text interrelation. On MUSIC-AVQA v2.0, this model surpasses all the existing benchmarks, improving accuracy by 2% on MUSIC-AVQA v2.0, setting a new state-of-the-art performance.

Reinforcement learning with verifiable rewards (RLVR) makes a language model solve more problems on the first try and fewer when given many tries: it raises pass@1 while the base model overtakes it at large sampling budget k. Is this a failure of RL to add reasoning, or a transient that more training removes? We settle the question with a theorem. The reversal is a price, and we compute its exact rate. Modeling RLVR as an exponential tilt of a base policy over strategies shared across prompts, we prove that under exclusive specialization every update that raises pass@1 strictly lowers the success of some prompt; a strictly positive price on prompts (Stiemke duality) is the exact condition, and being merely the best strategy on a prompt is provably not enough. When the loss reaches the hardest prompts, it forces a pass@k crossover, and the number of crossings is at most the sign variation of the shift in per-prompt success, a totally positive bound that is tight and that recovers the known single-crossover result as its order-preserving case. The crossover is not a bug to patch: a constrained update traces the exact pass@1 versus pass@k frontier, and removing the crossover entirely costs the entire pass@1 gain. The same tilt identity yields entropy collapse and the entropy-performance law with an identifiable exponent, unifying two literatures under one mechanism. Reinforcement cannot widen the reasoning frontier; it can only move along it, at a price we make measurable by a diagnostic that reads the bound, the order-preserving test, and the price from any evaluation log. There is no free lunch, only a price.

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.

Recipes such as s1 and LIMO teach a model to reason with little data by showing it the same thousand or fewer worked solutions many times over. Judged when that training ends, the repetition looks harmless. But reasoning models are often trained again, and we find that repetition leaves their reasoning fragile to that next stage, even when the stage has nothing to do with reasoning. We fine-tuned Qwen3.5-9B-Base on its own correct solutions to competition math problems, either drilling a few hundred of them about eight times each or showing many more once; with the same amount of training, both solve about 95% of held-out problems. A single pass of ordinary instruction tuning leaves the once-trained model where it was, while the drilled one falls to 86.0%, and harsher later stages take it to 59.3% or below. A third model that visited the drilled problems just as often, with a new solution at every visit, was unharmed, so the damage comes from seeing the same texts again rather than from having few problems. The break recurs with a stronger model's traces, in further training runs and on other models and tasks. It is also cheap to undo: the reasoning is suppressed rather than erased, and five updates of reasoning training bring almost all of it back, as does brief training on the reasoning format with almost no mathematics. Fresh solutions prevented the damage, and so did replaying 6.25% of the original solutions in a gentler later stage, so our claim concerns later training without such replay. Sharpening alone does not explain the break, since a model sharpened three-quarters as much without repetition was unharmed. On a skill the base model could not perform within a token budget, repetition mainly cost learning.

We present a new probing dataset named PROST: Physical Reasoning about Objects Through Space and Time. This dataset contains 18,736 multiple-choice questions made from 14 manually curated templates, covering 10 physical reasoning concepts. All questions are designed to probe both causal and masked language models in a zero-shot setting. We conduct an extensive analysis which demonstrates that state-of-the-art pretrained models are inadequate at physical reasoning: they are influenced by the order in which answer options are presented to them, they struggle when the superlative in a question is inverted (e.g., most <-> least), and increasing the amount of pretraining data and parameters only yields minimal improvements. These results provide support for the hypothesis that current pretrained models' ability to reason about physical interactions is inherently limited by a lack of real world experience. By highlighting these limitations, we hope to motivate the development of models with a human-like understanding of the physical world.

Large language models often produce hallucinated answers that violate prompt-level constraints. A key diagnostic question is whether these failures reflect missing knowledge, or whether the model has the relevant information but follows the wrong inference path. We study this phenomenon as inference misalignment: a mismatch between the answer supported by the prompt and the answer favored by statistically salient latent associations. We formalize this view with a latent key-task model, in which pretraining-frequency imbalance can cause a shortcut path to dominate the constraint-sensitive path and induce positive inference loss. The framework predicts two failure modes: task-retrieval bias in entity disambiguation and key-selection bias in action choice. We introduce TrapQA, a controlled diagnostic testbed with two components. ScientistQA tests disambiguation among similar scientists with supplementary factual probes, while Real-Life Constrained QA tests everyday constraint following under salient shortcuts. Our results show that hallucination can arise from biased latent inference rather than absent knowledge alone.

On-policy distillation (OPD) consolidates expert capabilities into a single student by matching teacher and student final-layer distributions on student-generated rollouts. We investigate whether depthwise prediction transitions, the changes in predictions between layers, can provide additional task-relevant supervision. Our diagnostic study finds that stronger teacher--student agreement in intermediate transitions is associated with higher answer correctness within questions, with correct responses exhibiting lower mean transition KL across all measured intermediate layer pairs. These observations motivate On-Policy Transition Distillation (OPTD), which augments final-layer distribution matching by aligning teacher and student layer-to-layer log-probability increments on student-generated rollouts. Evaluations across three backbones and eight benchmarks spanning mathematics, coding, and instruction following show that OPTD achieves the best or tied-best overall performance in both dual- and three-domain distillation. In the three-domain setting, OPTD recovers approx 90% of the overall teacher--student capability gap, compared with approx 75% for OPD.

If a question cannot be answered with the available information, robust systems for question answering (QA) should know _not_ to answer. One way to build QA models that do this is with additional training data comprised of unanswerable questions, created either by employing annotators or through automated methods for unanswerable question generation. To show that the model complexity of existing automated approaches is not justified, we examine a simpler data augmentation method for unanswerable question generation in English: performing antonym and entity swaps on answerable questions. Compared to the prior state-of-the-art, data generated with our training-free and lightweight strategy results in better models (+1.6 F1 points on SQuAD 2.0 data with BERT-large), and has higher human-judged relatedness and readability. We quantify the raw benefits of our approach compared to no augmentation across multiple encoder models, using different amounts of generated data, and also on TydiQA-MinSpan data (+9.3 F1 points with BERT-large). Our results establish swaps as a simple but strong baseline for future work.

Reverse Chain-of-Thought Generation (RCG) synthesizes reasoning traces from query-answer pairs, but runs the risk of producing post-hoc rationalizations: when models can see the answer during generation, the answer serves as a cognitive anchor that shapes the entire explanation. We formalize this phenomenon through a three-level measurement hierarchy: lexical, entropic, and probabilistic anchoring, each captures surface artifacts, entropy dynamics, and latent answer dependence, respectively. We analyze semantic suppression, the intuitive mitigation strategy that instructs models to ignore the answer, to find out its counterproduction: while it reduces lexical overlap, it paradoxically increases entropic and probabilistic anchoring. Drawing on Ironic Process Theory from cognitive psychology, we attribute this failure to active monitoring of the forbidden answer, which inadvertently deepens dependence on it. To break this cycle, we propose Structural Skeleton-guided Reasoning (SSR), a two-phase approach that first generates an answer-invariant functional skeleton structure, then uses this skeleton to guide full trace generation. By redirecting the information flow to structural planning rather than answer monitoring, SSR consistently reduces anchoring across all three levels. We further introduce Distilled SSR (SSR-D), which fine-tunes models on teacher-generated SSR traces to ensure reliable structural adherence. Experiments across open-ended reasoning benchmarks demonstrate that SSR-D achieves up to 10% improvement over suppression baselines while preserving out-of-distribution (OOD) generalization.

Large Language Models (LLMs) are expected to provide helpful and harmless responses, yet they often exhibit sycophancy--conforming to user beliefs regardless of factual accuracy or ethical soundness. Prior research on sycophancy has primarily focused on single-turn factual correctness, overlooking the dynamics of real-world interactions. In this work, we introduce SYCON Bench, a novel benchmark for evaluating sycophantic behavior in multi-turn, free-form conversational settings. Our benchmark measures how quickly a model conforms to the user (Turn of Flip) and how frequently it shifts its stance under sustained user pressure (Number of Flip). Applying SYCON Bench to 17 LLMs across three real-world scenarios, we find that sycophancy remains a prevalent failure mode. Our analysis shows that alignment tuning amplifies sycophantic behavior, whereas model scaling and reasoning optimization strengthen the model's ability to resist undesirable user views. Reasoning models generally outperform instruction-tuned models but often fail when they over-index on logical exposition instead of directly addressing the user's underlying beliefs. Finally, we evaluate four additional prompting strategies and demonstrate that adopting a third-person perspective reduces sycophancy by up to 63.8% in debate scenario. We release our code and data at https://github.com/JiseungHong/SYCON-Bench.

Recent advances in Large Reasoning Models (LRMs) have shown impressive capabilities in mathematical and logical reasoning. However, current LRMs rarely admit ignorance or respond with "I don't know". Instead, they often produce incorrect answers while showing undue confidence, raising concerns about their factual reliability. In this work, we identify two pathological reasoning patterns characterized by overthinking that contribute to the overconfident and incorrect answers: last-minute guessing and second-thought spiraling. To address these issues, we propose BARREL-a novel framework that promotes concise and boundary-aware factual reasoning. Our experiments show that BARREL-training increases the reliability of DeepSeek-R1-Distill-Llama-8B from 39.33% to 61.48%, while still achieving accuracy comparable to models finetuned on reasoning data generated by R1. These results demonstrate that our pilot study is inspiring to build more reliable and factual System 2 LRMs.

Long-horizon LLM agents depend on intermediate information-gathering turns, yet training feedback is usually observed only at the final answer, because process-level rewards require high-quality human annotation. Existing turn-level shaping methods reward turns that increase the likelihood of a gold answer, but they require answer supervision or stable task-specific verifiers. Conversely, label-free RL methods extract self-signals from output distributions, but mainly at the answer or trajectory level and therefore cannot assign credit to intermediate turns. We propose Self-Induced Outcome Potential (SIOP), which treats semantic clusters of final answers as latent future outcome states for potential-based turn-level credit assignment. For each query, SIOP samples multiple rollouts, clusters final answers into semantic outcome modes, and builds a reliability-aware target distribution over these states. It then rewards turns for increasing posterior support for reliable future states using a tractable cluster-level approximation. The objective generalizes information-potential shaping from gold-answer supervision to settings without task-specific gold verifiers while avoiding the broadcasted rollout-level advantages used by standard GRPO. We formalize the framework, characterize its supervised gold-answer limit, and show that SIOP improves average performance over verifier-free outcome-level baselines on seven search-augmented agentic reasoning benchmarks while approaching a gold-supervised outcome baseline. Code is available at https://github.com/dl-m9/SIOP.git.

Large language models often respond to ambiguous requests by implicitly committing to one interpretation. Intent misunderstandings can frustrate users and create safety risks. To address this, we propose generating multiple interpretation-answer pairs in a single structured response to ambiguous requests. Our models are trained with reinforcement learning and customized reward functions using multiple valid answers as supervision. Experiments on conversational question answering and semantic parsing demonstrate that our method achieves higher coverage of valid answers than baseline approaches. Human evaluation confirms that predicted interpretations are highly aligned with their answers. Our approach promotes transparency with explicit interpretations, achieves efficiency by requiring only one generation step, and supports downstream applications through its structured output format.

Sycophancy is a language model's tendency to cave when a user pushes back, abandoning a correct answer for the user's. Several benchmarks now measure it by scripting an objection and recording how often the model caves. Because that objection is a prompt template, whatever else the template varies is measured along with the property it claims to isolate. We audit SycEval, which reports that objections raised before a model answers (preemptive) cause more caving than those raised after (in-context), and attributes the gap to timing. Two features of its templates vary alongside the property each is meant to test. First, SycEval's objections escalate through four strength levels, and at the two weakest only the preemptive template names a target answer, so timing and naming vary together. We build the missing comparison and test it on multiple-choice questions and SycEval's own free-form pipeline. Naming a target answer raises the follow rate, the share of samples matching the user's assertion, by 14.1--49.5 percentage points (pp), and once both templates name one, the timing comparison reverses on three of the five model conditions we test: models cave less under preemptive objections than in-context ones, opposite to SycEval. Second, the output-format instruction benchmarks append for automatic grading also varies with objection placement. Moving it from the pushback into the question flips its effect in opposite directions across models on the same items (p=0.0059). The same effect appears in SycEval's free-form pipeline: relocating its instruction lowers caving by 5.1pp on Llama-3.1-8B, while an equivalence test confirms no effect on Qwen3-4B. Both confounds live in the template rather than the models under test, so both are correctable: we close with three checks benchmark authors can apply before publishing.

Bias scores in language-model evaluations often conflate what a model prefers with whether it answers at all. We identify this confound in BBQ, whose ambiguous score factorizes into answerability and signed preference. Across 11 categories, 28,}824 items, and five models, contrastive steering lowers the signed ambiguous score in 11/11 categories but lowers the signed disambiguated preference score in only 8/11. Across 99 configurations, the answerability channel exceeds half of the observed score change in 17\% and accounts for the entire change in 11%. In a six-category disambiguated evaluation, 20.8\% of 1,500 items flip to the unknown option, where abstention is incorrect; StereoSet shows no analogous answer-rate shift. The effect reflects selection and near-orthogonality between the contrastive and preference-gradient directions. We introduce Gradient-Constrained Direction Design (GCD), an option-gradient construction with a first-order unknown-logit constraint. Under a matched Qwen-7B protocol, GCD reduces answer-rate distortion 14-fold and yields a 4-fold larger preference change, while ablations show similar gains for alternative and unconstrained option-gradient directions; nulling the unknown gradient alone is not responsible for the gain. We recommend reporting answerability alongside published scores.

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.

We post-train Qwen3.8-27B for Korean response style -- verbosity, list and markdown usage, discourse structure and register -- and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text -- observable without any reference labels.

We introduce an evaluation methodology for visual question answering (VQA) to better diagnose cases of shortcut learning. These cases happen when a model exploits spurious statistical regularities to produce correct answers but does not actually deploy the desired behavior. There is a need to identify possible shortcuts in a dataset and assess their use before deploying a model in the real world. The research community in VQA has focused exclusively on question-based shortcuts, where a model might, for example, answer "What is the color of the sky" with "blue" by relying mostly on the question-conditional training prior and give little weight to visual evidence. We go a step further and consider multimodal shortcuts that involve both questions and images. We first identify potential shortcuts in the popular VQA v2 training set by mining trivial predictive rules such as co-occurrences of words and visual elements. We then introduce VQA-CounterExamples (VQA-CE), an evaluation protocol based on our subset of CounterExamples i.e. image-question-answer triplets where our rules lead to incorrect answers. We use this new evaluation in a large-scale study of existing approaches for VQA. We demonstrate that even state-of-the-art models perform poorly and that existing techniques to reduce biases are largely ineffective in this context. Our findings suggest that past work on question-based biases in VQA has only addressed one facet of a complex issue. The code for our method is available at https://github.com/cdancette/detect-shortcuts.

Full-duplex speech models listen and speak at once, promising always-on assistants. Yet they must also decide when they should speak. Human listeners speak when addressed or when the speaker stops, but also self-select to correct a false claim, supply a missing word, or warn of danger. We ask whether full-duplex models do the same. To separate the reason to speak from the opportunity, we construct context-matched English monologues in which only the trigger utterance varies within a topic, define 10 conditions from turn-allocation rules, and compress inter-word pauses to limit opportunities created by silence. Across five model families, being addressed and silence are far more reliable triggers than false facts or hazards. Frame-level text-token probabilities in Moshi and PersonaPlex are lower for false facts than for Neutral when averaged over the first 2\,s after trigger end. Pauses or permission to interrupt do not close this gap either. Given the floor, Moshi and PersonaPlex answer most direct questions, yet the proportion of non-empty false-fact replies that challenge the claim is only .14--.15, and the proportion of hazard replies that warn of danger is .04--.07. This paper thus identifies a gap in both speech initiation and response content. Closing it requires genuine content understanding and intervention decisions grounded in it.

Streaming omni-modal models must decide what and when to answer from the video chunks and synchronized audio observed so far. Visual cues often support an interpretation before an utterance or sound event is complete. If that interpretation enters memory as a fact, later reasoning can keep relaying it even after audio contradicts it. We call this failure premature cross-modal commitment. We propose Omni-Streaming Thinking (OST), which generates structured outputs that include evidence observed so far, forecasts of future evidence, and claims based on this evidence. Each claim is initially marked as pending and linked to a future verification interval. Audio and visual evidence are stored separately, and OST checks a claim against the evidence from the specified modality at the end of the verification interval. When contradictory evidence is detected, a refutation process reduces the influence of the claim and its dependent states, and then guides a state update using the new evidence. An answer gate decides whether the answer-critical claims meet the conditions for giving a response. Using a frozen Qwen3-Omni-30B-A3B-Instruct backbone with lightweight adaptation, OST outperforms the strongest open baselines on five streaming and audio-visual benchmarks by more than 10% relative on average. We also introduce OST-DiagBench, which holds video fixed and edits audio to test agreement, absence, contradiction, coexistence, and subtitle-speech conflict. OST reaches d-prime = 2.95, compared with at most 1.38 for open baselines, while reducing vision-induced auditory hallucinations.

Reverse thinking plays a crucial role in human reasoning. Humans can reason not only from a problem to a solution but also in reverse, i.e., start from the solution and reason towards the problem. This often enhances overall reasoning performance as it enables consistency checks between their forward and backward thinking. To enable Large Language Models (LLMs) to perform reverse thinking, we introduce Reverse-Enhanced Thinking (RevThink), a framework composed of data augmentation and learning objectives. In RevThink, we augment the dataset by collecting structured forward-backward reasoning from a teacher model, consisting of: (1) the original question, (2) forward reasoning, (3) backward question, and (4) backward reasoning. We then employ three objectives to train a smaller student model in a multi-task learning fashion: (a) generate forward reasoning from a question, (b) generate a backward question from a question, and (c) generate backward reasoning from the backward question. Experiments across 12 datasets covering commonsense, math, and logical reasoning show an average 13.53% improvement over the student model's zero-shot performance and a 6.84% improvement over the strongest knowledge distillation baselines. Moreover, our method demonstrates sample efficiency -- using only 10% of the correct forward reasoning from the training data, it outperforms a standard fine-tuning method trained on 10x more forward reasoning. RevThink also exhibits strong generalization to out-of-distribution held-out datasets.

Vision-language systems combine images with retrieved text, but these sources can disagree or jointly fail to support an answer. Reliable models must identify the trustworthy source and abstain when neither is adequate. Existing post-training objectives score instances independently and therefore do not enforce coherent behavior under counterfactual evidence changes. We introduce CARGO-VL, a group-relative framework that optimizes matched variants covering aligned, image-correct, text-correct, and both-wrong (A/V/T/N) evidence states as one bundle. Its objective couples condition-wise correctness with transition rewards for answer invariance, source equivariance, and answer-to-abstention switching, while a primal-dual controller balances unsafe answers against excessive deferral. We also contribute XMC (eXtended Modal Conflict), a four-condition conflict training resource, and evaluate transfer on CMC-Bench and Modality-Bias. Across multiple seeds, CARGO-VL improves conflict handling, unsupported-answer avoidance, and modality balance over pointwise baselines. Ablations identify complementary benefits from relational transition signals and adaptive risk control, supporting counterfactual consistency as a practical objective for reliable multimodal evidence arbitration.

Medical reasoning models (MRMs) achieve superior performance on medical benchmarks compared to medical LLMs; however, high accuracy alone is insufficient for practical deployment. One of such requirements for real-world application is robustness to varying output constraints. Specifically, posing the same medical question while requesting different answer formats should not affect the underlying correctness of the response. We investigate this phenomenon in this paper, focusing on MRMs. To quantify this behavior, we propose the metric answer-format robustness: the ability to reliably generate correct outputs across varying specified formats. We examine three representative formats: multiple-choice, open-ended question-answering, and ranked lists. Across 15 proprietary and open-weight models, we observe substantial variation in format robustness (35-100%). Furthermore, we conduct controlled fine-tuning experiments on a shared backbone with matched training data to isolate the effects of the fine-tuning paradigm. We find that supervised fine-tuning yields more stable behavior across formats, whereas reinforcement fine-tuning often exhibits higher cross-format brittleness, with the degree of instability strongly dependent on reward design. Overall, answer-format robustness in MRMs is trainable yet brittle and requires careful evaluation for practical medical use.

Many questions about a video concern one physical object. Where does the ball end up after the cups are shuffled? Which card is the queen? How often was the mug covered? Video-language models (Video-LMs) recognise such objects in single frames but do not keep them apart over time. On real shell-game videos, open models from 8B to 235B parameters, their thinking variants, a reinforcement-learned tool agent and three training-free tool agents all stay within ten points of chance. We separate semantic identity binding from temporal identity maintenance. An object's identity should be fixed at a moment when the evidence for it is visible, and carried afterwards by a tracker. Track, Then Bind implements this principle without training. A promptable video tracker follows every candidate object and draws it with a track number. A frozen vision-language model answers a static question about the identifying evidence, which binds the target to a track. The answer is then derived from the tracked state in text, without asking the model to reinterpret the original video during read-out. A single configuration covers all benchmarks. With Qwen3-VL-32B it solves both VET-Bench tasks, where a fine-tuned trajectory model reaches 92\% and 90\% and the same VLM voting over 16 samples stays at 30--32\%. It answers 55.4\% of the hardest Perception Test cup games, where the best published model answers 41.5\%, and, using track-derived timelines, 48.6\% of TOC-Bench questions, against 37.1\% for the same model answering directly. Drawing the same tracks on the video and asking the question directly falls 12--19 points short of the agent on real cup games.

System One models answer typed questions in a single forward pass and promise the same answer for the same input. For choice questions, this promise fails: the answer depends on the order of the options, increasingly so as they grow in number, and averaging over their cyclic shifts reduces this dependence without removing it. Scoring each option alone, which the order cannot affect, shows that the order changes the answer mostly when the model is unsure among its top few options, even though the listwise answer looks confident. We therefore propose Runoff, which considers, then chooses: it scores each option alone to keep a few candidates, then compares them in listings where each candidate takes every position. It needs no training and can run in two forward passes, and its answer does not depend on the option order by construction. On two open models and five datasets, it changes no answer under shuffling, against 9.3% for cyclic averaging, and with 20 options it is about as accurate and three to four times faster. Wrapped around eight System One models, it removes the order dependence of the five benchmarks, while cyclic averaging changes up to 15.4%.

Knowledge distillation aims to transfer the factual knowledge of large language models to smaller models for efficient deployment. Yet a teacher may recall a relation in one direction while failing to generate the answer in the reverse direction. Distillation from its generated answers can therefore propagate this directional limitation to the student. The same teacher can nevertheless recognize such an answer by scoring the relation in the direction it knows. We introduce directional label distillation, in which frozen teachers score candidate answers in that known direction and the best-scoring candidate becomes the student's training target. On facts about parents and their children, known-direction scoring yields more accurate labels than scoring the requested direction, even after tuned corrections for name priors. With prior-corrected scores, the better direction depends on the facts rather than the template, and reverses on mined facts whose notable entity is the parent rather than the child. With the evaluated children's forward facts withheld, students trained on known-direction labels improve open-ended accuracy on their trained queries by 13 to 15 points over students trained on prior-corrected reverse labels. After generated answers are matched to a fixed name list by lexical similarity, students reproduce nearly all selected labels. Their accuracy largely follows label quality. The label advantage holds on unscreened queries and when candidates are retrieved without inserting correct answers. Our findings show that directional verification mitigates the transfer of errors from teacher-generated answers to students by providing more accurate training targets. Code is available at https://github.com/js-lee-AI/directional-verification.

An ordinary-looking background detail can turn a correct model decision into a confident mistake. We demonstrate this fragility in four decision systems, including Jev, across seven datasets covering knowledge, reasoning, and tool routing. Within 64 accepted target evaluations per decision, we uncover short context additions that redirect 61.4%-73.2% of each system's initially correct decisions toward a wrong option fixed in advance. The additions supply background or procedural information rather than explicit answer-selection instructions, leaving the original question and choices intact. We construct them through probability-guided context optimization, which uses shifts in the option distribution to refine surrounding text under naturalness and answer-preservation constraints. Redirection affects initially confident decisions, often produces high-confidence wrong choices, and transfers across models. In blinded human evaluation, 91.6% of 250 sampled successful contexts are judged natural, answer-preserving, and free of decisive answer-changing evidence by a majority of three independent annotators. These findings expose a weakness in current decision models: context that looks entirely compatible with an input can redirect the choices that agents, routers, and evaluators rely on.

We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form "A is B", it will not automatically generalize to the reverse direction "B is A". This is the Reversal Curse. For instance, if a model is trained on "Olaf Scholz was the ninth Chancellor of Germany", it will not automatically be able to answer the question, "Who was the ninth Chancellor of Germany?". Moreover, the likelihood of the correct answer ("Olaf Scholz") will not be higher than for a random name. Thus, models exhibit a basic failure of logical deduction and do not generalize a prevalent pattern in their training set (i.e. if "A is B'' occurs, "B is A" is more likely to occur). We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as "Uriah Hawthorne is the composer of 'Abyssal Melodies'" and showing that they fail to correctly answer "Who composed 'Abyssal Melodies?'". The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as "Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]" and the reverse "Who is Mary Lee Pfeiffer's son?". GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter. This shows a failure of logical deduction that we hypothesize is caused by the Reversal Curse. Code is available at https://github.com/lukasberglund/reversal_curse.

We employ an inversion-based approach to examine CLIP models. Our examination reveals that inverting CLIP models results in the generation of images that exhibit semantic alignment with the specified target prompts. We leverage these inverted images to gain insights into various aspects of CLIP models, such as their ability to blend concepts and inclusion of gender biases. We notably observe instances of NSFW (Not Safe For Work) images during model inversion. This phenomenon occurs even for semantically innocuous prompts, like "a beautiful landscape," as well as for prompts involving the names of celebrities.

One of the most widely used methods to evaluate LLMs are Multiple Choice Question (MCQ) tests. MCQ benchmarks enable the testing of LLM knowledge on almost any topic at scale as the results can be processed automatically. To help the LLM answer, a few examples called few shots can be included in the prompt. Moreover, the LLM can be asked to answer the question directly with the selected option or to first provide the reasoning and then the selected answer, which is known as chain of thought. In addition to checking whether the selected answer is correct, the evaluation can look at the LLM-estimated probability of its response as an indication of the confidence of the LLM in the response. In this paper, we study how the LLM confidence in its answer depends on whether the model has been asked to answer directly or to provide the reasoning before answering. The results of the evaluation of questions on a wide range of topics in seven different models show that LLMs are more confident in their answers when they provide reasoning before the answer. This occurs regardless of whether the selected answer is correct. Our hypothesis is that this behavior is due to the reasoning that modifies the probability of the selected answer, as the LLM predicts the answer based on the input question and the reasoning that supports the selection made. Therefore, LLM estimated probabilities seem to have intrinsic limitations that should be understood in order to use them in evaluation procedures. Interestingly, the same behavior has been observed in humans, for whom explaining an answer increases confidence in its correctness.

Ask a language model the same question about the same document twenty times, and it sometimes returns two different answers. We built Probity, a benchmark of 60 tasks and 470 items from real venture-financing filings, to measure how often this happens. Then we audited our own corpus and found a defect any excerpt-built benchmark can carry: items whose evidence is missing from the window of text the model is shown. The audit flags 36 items and separates two failures a single flag would conflate: evidence genuinely absent from the window and answers that must be computed from numbers the window does supply. Flagged items change their answers far more often, wobbling at 0.255 against 0.087 on the 427 clean items, and excluding them cuts apparent cross-model agreement by about a fifth. Before testing whether the missing evidence explains the instability, we registered a prediction: re-cut each window to hold its evidence, and instability should fall below a set threshold. It failed: the repair moved wobble by 0.058, with an interval containing zero. We report the association as correlational. Almost all measurements sit where instability cannot show, which bounds what a corpus built for accuracy can say about stability. We release the corpus, all 112,800 raw responses, and the audit as a runnable check for any document benchmark.

Recent advancements in reinforcement learning, particularly through Group Relative Policy Optimization (GRPO), have significantly improved multimodal large language models for complex reasoning tasks. However, two critical limitations persist: 1) they often produce unfocused, verbose reasoning chains that obscure salient spatiotemporal cues and 2) binary rewarding fails to account for partially correct answers, resulting in high reward variance and inefficient learning. In this paper, we propose TW-GRPO, a novel framework that enhances visual reasoning with focused thinking and dense reward granularity. Specifically, we employs a token weighting mechanism that prioritizes tokens with high informational density (estimated by intra-group variance), suppressing redundant tokens like generic reasoning prefixes. Furthermore, we reformulate RL training by shifting from single-choice to multi-choice QA tasks, where soft rewards enable finer-grained gradient estimation by distinguishing partial correctness. Additionally, we propose question-answer inversion, a data augmentation strategy to generate diverse multi-choice samples from existing benchmarks. Experiments demonstrate state-of-the-art performance on several video reasoning and general understanding benchmarks. Notably, TW-GRPO achieves 50.4\% accuracy on CLEVRER (18.8\% improvement over Video-R1) and 65.8\% on MMVU. Our codes are available at https://github.com/longmalongma/TW-GRPO.

Language models learn from web documents, some of them false, and false content can reach a model's answer to a factual question and the summaries and decisions that use it. Most data-poisoning studies add a trigger to the training data and activate it in the prompt. False documents can also change factual responses without any trigger, but we do not know whether the direct answer predicts the decision. In this work, we follow false content past the answer and find an audit gap between what a direct probe reports and what the model then does, comparing false training with matched truthful controls in a controlled decision task, Guess the Capital, where a fixed decoder turns factual answers into a scored card choice, and on a misleading claim from Facebook posts about the 2019--20 Australian bushfires. Across eight models at dose 1,000, direct injected-choice rates reach 95.8--100\%, while injected game choices increase by 1.7--14.4 percentage points over matched truthful training. The gap runs the other way too. Facts that pass the direct probe still push decisions toward the injected answer, and game accuracy drops further than those choices explain. Truthful correction brings the fact back but not the decisions built on it. We then look into the real-world bushfire case, models trained on the false posts say that people were arrested for arson even when they lose the inflated count, and in a count-by-wording factorial the misleading arrest wording produces arrest assertions even when the training count stays at 24. In simpler terms, a correct factual answer does not guarantee a correct decision, and losing the injected number does not remove the misleading story.

System One models choose an agent's next action by scoring a supplied set of options in a single pass. This makes them fast and cheap, but when they choose wrongly there is nothing to inspect: did the model misread what the agent has seen, or read it correctly and still act wrongly? We answer behaviorally. On a frozen benchmark of agent decision states from simulated environments, customer-service policies, tool use, web navigation and coding, we pair every next-action question with factual probes about the same state, each answered independently. Across seventeen fast systems and a reasoning model, the answer is consistent: most wrong actions follow correct reading, and the more accurate a system, the more of its errors take this form. These errors are not noise; they gather on the same states across systems of every kind. Where the probed facts provably determine the action, models that read every fact still fail to combine them, and their own answers, composed by code, repair the failure. On real states, what is missing is often not a fact but the next sub-goal.

When a speaker says ``some students passed'', listeners infer that not all of them did. Reasoning models now let users decide how long a model thinks before it answers, but we do not know what extra thinking does to such everyday inferences. We study this directly: we keep a model's weights fixed, change only how much it thinks, and record which reading it chooses. A second tool cuts the model's own reasoning at successive points and asks for an answer, showing where the reading changes. Thinking makes models read logical words as logic: it cancels the implicatures of some (not all) and or (not both) in most tested open reasoning models and in GPT-5, while context-driven inferences such as indirect answers improve. The trace shows why: models first entertain the enriched reading and cancel it later. Answering from that early reading, an exploratory inference-time remedy, recovers these implicatures without losing literal accuracy. Evaluations of reasoning models should measure how they interpret, not only how accurate they are.

This paper measures hour-of-day effects, filter precision, fragility, and realised yield in a 15-day paper-traded deployment of an autonomous memecoin trading system on Solana decentralised exchanges. The 190-trade sample (March 29 to April 12, 2026) shows a 40.5 percent win rate, mean per-trade return of +0.62 percent, cumulative +117.7 percent (net SOL +0.039), skewness -1.21, excess kurtosis 6.61. A Mann-Whitney U test of three poorest-performing UTC hours (2, 13, 23) against the others yields U = 1,274, p = 0.22; directional but not significant at n = 190. The three hours were selected in-sample, so the comparison is exploratory, not confirmatory. A parallel counterfactual rejection-tracking system collected 4,874 forward-sample observations across 184 distinct rejection events. Of those events, 17.9 percent reached a 50 percent drawdown from reference within 24 hours; 26.0 percent of forward samples recorded the rejected token below half-reference. The filter stack avoided these realised drawdowns: evidence that the rejection criteria are net-positive against forward-market outcomes. Fragility is the principal caveat. Removing the top three trades (1.6 percent of sample) flips cumulative return unprofitable. Profitability rests on a small number of large winners and is structurally fragile. The dataset and audit script are deposited under CC-BY-4.0 (Zenodo DOI 10.5281/zenodo.20043302).

Visual question answering is a task of predicting the answer to a question about an image. Given that different people can provide different answers to a visual question, we aim to better understand why with answer groundings. We introduce the first dataset that visually grounds each unique answer to each visual question, which we call VQAAnswerTherapy. We then propose two novel problems of predicting whether a visual question has a single answer grounding and localizing all answer groundings. We benchmark modern algorithms for these novel problems to show where they succeed and struggle. The dataset and evaluation server can be found publicly at https://vizwiz.org/tasks-and-datasets/vqa-answer-therapy/.

Large language models can answer scientific questions, yet a correct output does not reveal whether the model represents or uses the governing physics. Here we show that materials science mechanism information in the open-weight google/gemma-4-E4B-it model has three experimentally separable forms: concepts are readable in individual hidden states, constitutive orientation is carried by controlled transformations between states, and selected internal representations causally control engineering answers. We combine matched direct and Jacobian vocabulary readouts, option-free state geometry, a 60-law counterfactual benchmark and causal interventions. In 50 held-out materials descriptions, three independently fitted Jacobian lenses reproduced concept ranks, and target-free word sets from both readouts enabled blinded identification of 9 of 10 mechanism families. A separate 72-prompt benchmark produced mechanism-specific hidden-state neighborhoods, but an exact graph audit showed that this apparent physical organization was equally explained by numerical comparison. We therefore compared otherwise identical prompts in which only the direction of the physical input was reversed, asking whether the resulting hidden-state movement followed the supplied constitutive law. These state transformations ordered direct, physically neutral and inverse laws across 60 frozen relations and correctly oriented 39 of 40 directional laws, whereas lexical controls were near chance. Bidirectional interventions shifted answer probabilities toward or away from the physically appropriate outcome across all 12 matched cases, while counterfactual state patches transferred opposing decision signals across mechanisms and answer formats. Physical relationships were therefore more visible in controlled state changes than in absolute states alone.

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

Large reasoning models (LRMs) show strong capabilities in complex reasoning, yet their marginal gains on evidence-dependent factual questions are limited. We find this limitation is partially attributable to a reasoning-answer hit gap, where the model identifies the correct facts during reasoning but fails to incorporate them into the final response, thereby reducing factual fidelity. To address this issue, we propose MR-ALIGN, a Meta-Reasoning informed alignment framework that enhances factuality without relying on external verifiers. MR-ALIGN quantifies state transition probabilities along the model's thinking process and constructs a transition-aware implicit reward that reinforces beneficial reasoning patterns while suppressing defective ones at the atomic thinking segments. This re-weighting reshapes token-level signals into probability-aware segment scores, encouraging coherent reasoning trajectories that are more conducive to factual correctness. Empirical evaluations across four factual QA datasets and one long-form factuality benchmark show that MR-ALIGN consistently improves accuracy and truthfulness while reducing misleading reasoning. These results highlight that aligning the reasoning process itself, rather than merely the outputs, is pivotal for advancing factuality in LRMs.

Most existing multi-hop datasets are extractive answer datasets, where the answers to the questions can be extracted directly from the provided context. This often leads models to use heuristics or shortcuts instead of performing true multi-hop reasoning. In this paper, we propose a new multi-hop dataset, MoreHopQA, which shifts from extractive to generative answers. Our dataset is created by utilizing three existing multi-hop datasets: HotpotQA, 2WikiMultihopQA, and MuSiQue. Instead of relying solely on factual reasoning, we enhance the existing multi-hop questions by adding another layer of questioning that involves one, two, or all three of the following types of reasoning: commonsense, arithmetic, and symbolic. Our dataset is created through a semi-automated process, resulting in a dataset with 1,118 samples that have undergone human verification. We then use our dataset to evaluate five different large language models: Mistral 7B, Gemma 7B, Llama 3 (8B and 70B), and GPT-4. We also design various cases to analyze the reasoning steps in the question-answering process. Our results show that models perform well on initial multi-hop questions but struggle with our extended questions, indicating that our dataset is more challenging than previous ones. Our analysis of question decomposition reveals that although models can correctly answer questions, only a portion - 38.7% for GPT-4 and 33.4% for Llama3-70B - achieve perfect reasoning, where all corresponding sub-questions are answered correctly. Evaluation code and data are available at https://github.com/Alab-NII/morehopqa

Language model agents are increasingly used to simulate social interaction, and the resulting transcripts read as though the agents understand one another. We ask whether that appearance rests on a model of the partner's mind or on the surface record of what the partner said. We build a social simulation in which both questions have exact answers: 40 multi-issue negotiations whose hidden preference weights and whose full Pareto frontier are known by construction. Two model families negotiate across 160 dyads, every transcript is frozen before any measurement, and 2880 counterfactual probes then hold the evidence byte identical while moving one factor at a time: the reader's own stake, the partner's tone, an identity label, and the order of recursion. The agents are socially fluent and economically poor. They reach agreement in 96.2% of dyads with 0 protocol failures, yet only 0.7% of deals land on the Pareto frontier, they leave 20.5% of the available joint value unclaimed, and they miss the one issue on which their interests are perfectly aligned in 76.6% of deals; on the frontier and on that aligned issue, a package drawn at random from the set both sides would accept does as well. The probes locate the failure. Swapping only the reader's own payoff sheet, while the partner's words and offers stay identical, moves the inferred top priority by 15.0 percentage points, which is egocentric projection rather than inference, while a tone rewrite moves it by 5.3 percentage points and an identity label by 0.0. Most tellingly, an agent predicts what its partner believes about it 72.5% of the time while that partner's belief is itself correct only 51.2% of the time: the agents track the conversation far better than they track the mind behind it.

Corruption studies, the primary tool for evaluating chain-of-thought (CoT) faithfulness, identify which chain positions are "computationally important" by measuring accuracy when steps are replaced with errors. We identify a systematic confound: for chains with explicit terminal answer statements, the dominant format in standard benchmarks, corruption studies detect where the answer text appears, not where computation occurs. A within-dataset format ablation provides the key evidence: on standard GSM8K chains ending with "the answer is X," removing only the answer statement, preserving all reasoning, collapses suffix sensitivity ~19x at 3B (N=300, p=0.022). Conflicting-answer experiments quantify the causal mechanism: at 7B, CC accuracy drops to near-zero (<=0.02) across five architecture families; the followed-wrong rate spans 0.63-1.00 at 3B-7B and attenuates at larger scales (0.300 at Phi-4-14B, ~0.01 at 32B). A within-stable 7B replication (9.3x attenuation, N=76, p=7.8e-3; Qwen3-8B N=299, p=0.004) provides converging evidence, and the pattern replicates on MATH (DeepSeek-R1-7B: 10.9x suffix-survival recovery). On chains without answer suffixes the same protocol identifies the prefix as load-bearing (Delta=-0.77, p<10^-12). Generation-time probes confirm a dissociation: the answer is not early-determined during generation (early commitment <5%), yet at consumption time model outputs systematically follow the explicit answer text. The format-determination effect persists through 14B (8.5x ratio, p=0.001) and converges toward zero at 32B. We propose a three-prerequisite protocol (question-only control, format characterization, all-position sweep) as a minimum standard for corruption-based faithfulness studies.

Do reasoning models have "Aha!" moments? Prior work suggests that models like DeepSeek-R1-Zero undergo sudden mid-trace realizations that lead to accurate outputs, implying an intrinsic capacity for self-correction. Yet, it remains unclear whether such intrinsic shifts in reasoning strategy actually improve performance. Here, we study mid-reasoning shifts and instrument training runs to detect them. Our analysis spans 1M+ reasoning traces, hundreds of training checkpoints, three reasoning domains, and multiple decoding temperatures and model architectures. We find that reasoning shifts are rare, do not become more frequent with training, and seldom improve accuracy, indicating that they do not correspond to prior perceptions of model insight. However, their effect varies with model uncertainty. Building on this finding, we show that artificially triggering extrinsic shifts under high entropy reliably improves accuracy. Our results show that mid-reasoning shifts are symptoms of unstable inference behavior rather than an intrinsic mechanism for self-correction.

When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.

Current question-answering benchmarks predominantly focus on accuracy in realizable prediction tasks. Conditioned on a question and answer-key, does the most likely token match the ground truth? Such benchmarks necessarily fail to evaluate LLMs' ability to quantify ground-truth outcome uncertainty. In this work, we focus on the use of LLMs as risk scores for unrealizable prediction tasks. We introduce folktexts, a software package to systematically generate risk scores using LLMs, and evaluate them against US Census data products. A flexible API enables the use of different prompting schemes, local or web-hosted models, and diverse census columns that can be used to compose custom prediction tasks. We evaluate 17 recent LLMs across five proposed benchmark tasks. We find that zero-shot risk scores produced by multiple-choice question-answering have high predictive signal but are widely miscalibrated. Base models consistently overestimate outcome uncertainty, while instruction-tuned models underestimate uncertainty and produce over-confident risk scores. In fact, instruction-tuning polarizes answer distribution regardless of true underlying data uncertainty. This reveals a general inability of instruction-tuned LLMs to express data uncertainty using multiple-choice answers. A separate experiment using verbalized chat-style risk queries yields substantially improved calibration across instruction-tuned models. These differences in ability to quantify data uncertainty cannot be revealed in realizable settings, and highlight a blind-spot in the current evaluation ecosystem that folktexts covers.

Large language models (LLMs) can correctly answer "When was Einstein born?" yet fail to provide the same date when writing about Einstein's life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs' answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs' trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.

We introduce Self-Verified Reasoner (SVR-R1), a multi-turn RL framework that turns a model's own verification into a learning signal for multimodal reasoning. For each query, the model proposes an answer using the same weights, and issues a binary self-verdict (Yes/No). A 'No' triggers a second-chance rethink; a 'Yes,' or a turn cap, finalizes the output for computing the outcome-based reward. SVR-R1 is implemented with GRPO and an asynchronous multi-turn rollout framework and needs no external supervision or auxiliary critics. We evaluate SVR-R1 on vision-language reasoning benchmarks and show that it improves accuracy by a large margin over strong standard GRPO baselines. Training dynamics show decreasing reliance on verification-fewer verification turns, yet higher test accuracy-indicating that the gap between verification and generation narrows as the policy internalizes self-correction and chooses the most confident answer via our framework. SVR-R1 bridges the less explored intersection of inference-time self-refinement and RL training for VLMs, offering a simple yet effective recipe for bootstrapping multimodal reasoning. We will open-source SVR-R1 to facilitate future research in VLMs.

Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a pre-registered synthetic probe, truncating the context window costs a low-cost reasoning model nothing in success when a needed fact can be looked up again (it looks the fact up instead), but drops its success to chance when the fact cannot be re-observed: the two penalties differ by 92.5 percentage points (90% CI [89.5, 95.5]). A pre-registered replication on one model from a second provider showed the same contrast. Once the fact was lost, that model chose no answer rather than guessing. Second, on ARC-AGI-3, where the benchmark's Standard harness trails its Provider Adapter harness by a wide margin, we decompose that gap one component at a time, for one model on four games. Summarizing old turns instead of dropping them, while keeping the Standard harness's invitation to take notes, recovers 86% of a 63-point gap without replaying the model's reasoning. The same harness without the invitation recovers 43%, and adding the provider's native replay of the model's reasoning recovers the remaining 57%. These are point estimates from one counted run per configuration and game, and the replay share depends on a single run. Third, in the benchmark's published runs and in ours, the Adapter's advantage in the probability of clearing a level it has reached is larger at deep levels than at early ones. This largely restates where runs stopped and does not separate a growing advantage from a constant one.

Recursive reasoning models such as the Tiny Recursive Model (TRM) iteratively refine two latent states: an answer state y, decoded into the output, and a reasoning state z, which is never decoded. Many of their failures have been attributed to bad basins, regions of latent space that trap the recursion on wrong answers. Recent test-time approaches perturb y, z or both to recover these failures, but the impact of this choice on performance, test-time compute efficiency and scaling properties remains unclear. Through controlled experiments on TRM, we find that a failed trajectory carries its failure in its answer state. At equal compute, in depth and in width, resampling y recovers more failures than any perturbation of z we tested, and the performance gap widens as the test-time compute budget grows. Our approach, Gated Answer Resampling (GAR), pairs this resampling of the answer state with the model's own halting head and restarts only the puzzles that the head rejects. With at most 64 restarts per puzzle and no training or hyperparameter search, GAR takes TRM from 85.9% to 99.4% on Sudoku-Extreme and from 68.8% to 85.6% on Pencil Puzzle Bench validation. On the same checkpoints, GAR also outperforms Probabilistic TRM (PTRM), which perturbs z at each step, even when PTRM uses 42 to 197 times as much compute.

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

Much recent work seeks to evaluate values and opinions in large language models (LLMs) using multiple-choice surveys and questionnaires. Most of this work is motivated by concerns around real-world LLM applications. For example, politically-biased LLMs may subtly influence society when they are used by millions of people. Such real-world concerns, however, stand in stark contrast to the artificiality of current evaluations: real users do not typically ask LLMs survey questions. Motivated by this discrepancy, we challenge the prevailing constrained evaluation paradigm for values and opinions in LLMs and explore more realistic unconstrained evaluations. As a case study, we focus on the popular Political Compass Test (PCT). In a systematic review, we find that most prior work using the PCT forces models to comply with the PCT's multiple-choice format. We show that models give substantively different answers when not forced; that answers change depending on how models are forced; and that answers lack paraphrase robustness. Then, we demonstrate that models give different answers yet again in a more realistic open-ended answer setting. We distill these findings into recommendations and open challenges in evaluating values and opinions in LLMs.

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).

Given a question, a language model (LM) implicitly encodes a distribution over possible answers. In practice, post-training procedures for LMs often collapse this distribution onto a single dominant mode. While this is generally not a problem for benchmark-style evaluations that assume one correct answer, many real-world tasks inherently involve multiple valid answers or irreducible uncertainty. Examples include medical diagnosis, ambiguous question answering, and settings with incomplete information. In these cases, we would like LMs to generate multiple plausible hypotheses, ideally with confidence estimates for each one, and without computationally intensive repeated sampling to generate non-modal answers. This paper describes a multi-answer reinforcement learning approach for training LMs to perform distributional reasoning over multiple answers during inference. We modify the RL objective to enable models to explicitly generate multiple candidate answers in a single forward pass, internalizing aspects of inference-time search into the model's generative process. Across question-answering, medical diagnostic, and coding benchmarks, we observe improved diversity, coverage, and set-level calibration scores compared to single answer trained baselines. Models trained with our approach require fewer tokens to generate multiple answers than competing approaches. On coding tasks, they are also substantially more accurate. These results position multi-answer RL as a principled and compute-efficient alternative to inference-time scaling procedures such as best-of-k. Code and more information can be found at https://multi-answer-rl.github.io/.