The Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial Pressure

• URL: http://arxiv.org/abs/2605.29087v1
• Date: Wed, 27 May 2026 20:41:08 GMT
• Title: The Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial Pressure
• Abstract summary: Reasoning models are evaluated on single-turn benchmarks but deployed in multi-turn dialogue, where users push back on correct answers.<n>We call this unfaithful capitulation (UC) and isolate it with a $2times 2$ latent-versus-behavioral framework that flip-rate metrics and single-turn probes both miss.
• Score: 8.86745721473138
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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 $2\times 2$ 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.

Related papers

• When Many Answers Are Valid, Voting Fails: Symbolic Verification for Best-of-K Causal Reasoning in LLMs [42.32694162421078]
  We introduce CALVER, a training-free symbolic verifier that scores structured traces against Pearl's causal criteria.<n>On CLEAR find-one-valid queries that admit multiple graph-valid answers, CALVER reaches 42.1% where plurality, a reward model, an LLM judge, and model confidence remain near 30%.
  arXiv  Detail & Related papers  (2026-08-04T11:45:46Z)
• Round-Trip Consistency: Bidirectional Diffusion Models Can Predict Their Own Rollout Errors [51.56484100374058]
  Autoregressive models accumulate error over long rollouts, yet at deployment there is no ground truth to measure it against.<n>We train a single conditional latent diffusion model that steps a dynamical system forward or backward in time via a direction flag.<n>We show that this bidirectionality supplies a measurement-free test-time error signal.
  arXiv  Detail & Related papers  (2026-08-01T13:49:46Z)
• Two Regimes of Chain-of-Thought Unfaithfulness: Behavioral Detection Fails Where Models Are Wrong [1.2744523252873352]
  Black-box (behavioral) detection of unfaithful CoT against FaithCoT-Bench's human annotations.<n>We find answer correctness structures the problem at every level.
  arXiv  Detail & Related papers  (2026-07-26T04:43:53Z)
• When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models [4.8190992438931035]
  LearnStop is a hidden-state-free logistic stopper over prefix-observable features.<n>A trajectory decomposition predicts the regime: learning pays where answers oscillate and correctness evidence is spread.<n>Cost accounting sharpens the picture further -- the same policy that saves 32% of tokens under KV-cache forking costs 121% extra under black-box repeated prefilling.
  arXiv  Detail & Related papers  (2026-06-29T19:33:42Z)
• Right Makes Might: Aligning Verified Hidden States Empowers RL Reasoning [55.264863369127774]
  Current methods reduce each correct rollout to a single reward bit, ignoring the geometric structure shared among their hidden states.<n>We propose Hidden-Align, an auxiliary loss function that aligns the last-layer hidden states of correct rollouts at the anchor token during RL training, with zero overhead in both training and inference.<n>On eight mathematical reasoning benchmarks, Hidden-Align improves average pass@1 over the DAPO baseline by 3.8, 6.2, and 5.4 percentage points on Qwen3-1.7B, 4B, and 14B respectively, with consistent pass@k gains across all three scales.
  arXiv  Detail & Related papers  (2026-06-02T06:51:15Z)
• Perception or Prejudice: Can MLLMs Go Beyond First Impressions of Personality? [74.69723255239663]
  Grounded Personality Reasoning requires MLLMs to anchor each Big Five rating in observable evidence through a chain of rating, reasoning, and grounding.<n> MM-OCEAN is produced by a multi-agent pipeline with human verification, with timestamped behavioral observations, evidence-grounded trait analyses, and seven categories of cue-grounding MCQs.<n>The analysis uncovers a striking Prejudice Gap: across the field, 51% of correct ratings are not grounded in retrieved cues, and the Holistic-Grounding Rate spans only 0-33.5%.
  arXiv  Detail & Related papers  (2026-05-21T07:42:47Z)
• CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization [50.59956036193097]
  A model produces a correct solution under reinforcement learning with verifiable rewards (RLVR)<n>Every token receives the same reward signal regardless of whether it was a decisive reasoning step or a grammatical filler.<n>We propose Contrastive Evidence Policy Optimization (CEPO)<n>CEPO asks a sharper question at every token: "does the correct answer favor this token?" but "does the correct answer favor it while the wrong answer disfavors it?"
  arXiv  Detail & Related papers  (2026-05-19T06:46:19Z)
• Drop the Act: Probe-Filtered RL for Faithful Chain-of-Thought Reasoning [1.3011345529764784]
  Reasoning models post-hoc rationalize answers they have already committed to internally.<n>ProFIL reduces post-commitment theater by **11--100%**, raises faithful-fraction (e.g., +24pp on LiveCodeBench under an independent Claude 3.7 Sonnet judge)<n>ProFIL also beats a matched length-penalty GRPO baseline, isolating the gain as semantic commitment-detection rather than chain compression.
  arXiv  Detail & Related papers  (2026-05-12T03:30:23Z)
• MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents [0.0]
  We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified evaluation framework.<n>We show faithful evaluation increases measured attack success by $4times$ (ASR-R: 0.25 to 1.00$)<n>Our primary contribution is MEMSAD, a calibration-based defense grounded in a gradient coupling.
  arXiv  Detail & Related papers  (2026-05-05T08:15:41Z)
• Just Keep Prompting: Evaluating Repetitive Socratic Prompting in VLMs [7.833541893955439]
  We introduce Just Keep Prompting (JKP), a multi-turn evaluation framework that measures VLM stability when users repeatedly challenge, question, or contradict a model's answer.<n>JKP probes models for up to 10 follow-up turns using three strategies: Adversarial Negation (repeated rejection), Pure Socratic Interrogation (repeated calls to reassess certainty), and Context-Aware Socratic Summarization.<n>We evaluate GPT-4o, Gemini 2.5 Pro, and Qwen3-VL-30B on a subset of the STAR benchmark across 720 multi-turn runs.
  arXiv  Detail & Related papers  (2026-05-01T17:02:19Z)
• One Token Away from Collapse: The Fragility of Instruction-Tuned Helpfulness [12.183451602438753]
  We show that simple lexical constraints (banning a single punctuation character or common word) cause instruction-tuned LLMs to collapse their responses.<n>We show that base models show no systematic collapse under identical constraints, with effects that are small, noisy, and bidirectional.
  arXiv  Detail & Related papers  (2026-04-14T17:40:01Z)
• Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought [11.955186033088351]
  We provide evidence of performative chain-of-thought (CoT) in reasoning models.<n>We compare activation probing, early forced answering, and a CoT monitor across two large models.<n>We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions.
  arXiv  Detail & Related papers  (2026-03-05T18:55:16Z)
• CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal [84.71254539482369]
  Group-relative reinforcement learning with verifiable rewards (RLVR) often wastes the most informative data it already has the failures.<n>We present CARE, a failure-centric post-training framework for multimodal reasoning that turns errors into supervision.<n> CARE improves accuracy and training smoothness while explicitly increasing the share of learning signal that comes from failures.
  arXiv  Detail & Related papers  (2025-12-22T16:34:21Z)
• RefineBench: Evaluating Refinement Capability of Language Models via Checklists [71.02281792867531]
  We evaluate two refinement modes: guided refinement and self-refinement.<n>In guided refinement, both proprietary LMs and large open-weight LMs can leverage targeted feedback to refine responses to near-perfect levels within five turns.<n>These findings suggest that frontier LMs require breakthroughs to self-refine their incorrect responses.
  arXiv  Detail & Related papers  (2025-11-27T07:20:52Z)
• Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning [46.106795445750855]
  REFRAIN is a training-free framework that determines when to stop reasoning to mitigate overthinking.<n> REFRAIN reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting.
  arXiv  Detail & Related papers  (2025-10-11T08:30:00Z)
• RESTRAIN: From Spurious Votes to Signals -- Self-Driven RL with Self-Penalization [52.01526898310723]
  We introduce RESTRAIN, a self-penalizing RL framework that converts the absence of gold labels into a useful learning signal.<n>Instead of overcommitting to spurious majority votes, RESTRAIN exploits signals from the model's entire answer distribution.<n>On challenging reasoning benchmarks, RESTRAIN delivers large gains using only unlabeled data.
  arXiv  Detail & Related papers  (2025-10-02T16:24:01Z)
• Faithful Chain-of-Thought Reasoning [51.21714389639417]
  Chain-of-Thought (CoT) prompting boosts Language Models' (LM) performance on a gamut of reasoning tasks. We propose Faithful CoT, a reasoning framework involving two stages: Translation and Problem Solving. This guarantees that the reasoning chain provides a faithful explanation of the final answer.
  arXiv  Detail & Related papers  (2023-01-31T03:04:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.