The Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial Pressure
On the map: LLM reasoning → Prompting and chain-of-thought reasoning
Abstract (arXiv)
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× 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 nothink -- 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.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 19 of its 44 references that are in the map. First posted on arXiv 27 May 2026.
Builds on
19 of 44 references.
Closest ancestors · 2
Same direction, newest first
- 2025Reasoning Models Don't Always Say What They Think8 May
- 2023Measuring Faithfulness in Chain-of-Thought Reasoning17 Jul
Other references · 17
- 2021Training Verifiers to Solve Math Word Problemscited by 10,109
- 2025Qwen3 Technical Reportcited by 8,820
- 2024GPT-4o System Cardcited by 4,588
- 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
- 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,986
- 2025gpt-oss-120b & gpt-oss-20b Model Cardcited by 1,386
- 2025s1: Simple test-time scalingcited by 1,341
- 2023Instruction-Following Evaluation for Large Language Modelscited by 1,228
- 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
- 2025LLMs Get Lost In Multi-Turn Conversationcited by 429
- 2024Judging the Judges: Evaluating Alignment and Vulnerabilities in LLMs-as-Judgescited by 205
- 2023Don't Make Your LLM an Evaluation Benchmark Cheatercited by 194
5 more
- 2023Simple synthetic data reduces sycophancy in large language modelscited by 187
- 2024A Survey on Recent Advances in LLM-Based Multi-turn Dialogue Systemscited by 170
- 2023Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experimentcited by 77
- 2023When Large Language Models contradict humans? Large Language Models' Sycophantic Behaviourcited by 62
- 2025Beyond Single-Turn: A Survey on Multi-Turn Interactions with Large Language Modelscited by 62
Built on by
Citing papers per year: 2026: 1.
1 paper.
Similar, not cited
Close in wording, no citation either way
- 2025Chain-of-Thought Reasoning In The Wild Is Not Always Faithful
- 2026Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models
- 2026When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel
- 2026Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models
- 2026When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models
- 2026Post-Hoc Reasoning in Chain of Thought: Decoding and Steering Pre-Committed Answers
- 2026Reasoning Concentrates Errors, and Self-Consistency Never Notices
- 2024Dissociation of Faithful and Unfaithful Reasoning in LLMs