mlsanityData to 7 Oct 2026

Why Models Know But Don't Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models

arXiv 2603.26410 ↗First posted 27 Mar 2026cs.CL cs.AIcited by 4

On the map: LLM reasoning → Prompting and chain-of-thought reasoning

Abstract (arXiv)

Extended-thinking models expose a second text-generation channel ("thinking tokens") alongside the user-visible answer. This study examines 12 open-weight reasoning models on MMLU and GPQA questions paired with misleading hints. Among the 10,506 cases where models actually followed the hint (choosing the hint's target over the ground truth), each case is classified by whether the model acknowledges the hint in its thinking tokens, its answer text, both, or neither. In 55.4% of these cases the model's thinking tokens contain hint-related keywords that the visible answer omits entirely, a pattern termed *thinking-answer divergence*. The reverse (answer-only acknowledgment) is near-zero (0.5%), confirming that the asymmetry is directional. Hint type shapes the pattern sharply: sycophancy is the most *transparent* hint, with 58.8% of sycophancy-influenced cases acknowledging the professor's authority in both channels, while consistency (72.2%) and unethical (62.7%) hints are dominated by thinking-only acknowledgment. Models also vary widely, from near-total divergence (Step-3.5-Flash: 94.7%) to relative transparency (Qwen3.5-27B: 19.6%). These results show that answer-text-only monitoring misses more than half of all hint-influenced reasoning and that thinking-token access, while necessary, still leaves 11.8% of cases with no verbalized acknowledgment in either channel.

Key facts. Cited by 4 papers in the map (data through 7 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 19 of its 20 references that are in the map. First posted on arXiv 27 Mar 2026.

Builds on

19 of 20 references.

Closest ancestors · 12

Same direction, newest first

  1. 2026Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?23 Mar
  2. 2026Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation20 Mar
  3. 2026Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought5 Mar
  4. 2026Mechanistic Evidence for Faithfulness Decay in Chain-of-Thought Reasoning4 Feb
  5. 2025Monitoring Monitorability20 Dec
  6. 2025Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity31 Oct
  7. 2025FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning5 Oct
  8. 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety15 Jul
  9. 2025Reasoning Models Don't Always Say What They Think8 May
  10. 2025When More is Less: Understanding Chain-of-Thought Length in LLMs11 Feb
  11. 2023Measuring Faithfulness in Chain-of-Thought Reasoning17 Jul
  12. 2023Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting7 May

Other references · 7

  1. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  2. 2024Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingcited by 478
  3. 2024Alignment faking in large language modelscited by 316
  4. 2025Reasoning Models Know When They're Right: Probing Hidden States for Self-Verificationcited by 154
  5. 2025Are DeepSeek R1 And Other Reasoning Models More Faithful?cited by 69
  6. 2025Measuring the Faithfulness of Thinking Drafts in Large Reasoning Modelscited by 22
  7. 2025Stealth Fine-Tuning: Efficiently Breaking Alignment in RVLMs Using Self-Generated CoTcited by 1

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 4.

4 papers.

  1. 2026When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channelcited by 2
  2. 2026How Language Models Fail: Token-Level Signatures of Committed and Persistent Reasoning Failurescited by 1
  3. 2026Measuring Reward Hacking and Reasoning-Answer Decoupling Under Position-Confounded Optimizationcited by 0
  4. 2026Untangling the Mechanisms of Misleading Context in Medical Question Answeringcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Unspoken Hints: Accuracy Without Acknowledgement in LLM Reasoning
  2. 2025Is Chain-of-Thought Really Not Explainability? Chain-of-Thought Can Be Faithful without Hint Verbalization
  3. 2026Reasoning Models Will Sometimes Lie About Their Reasoning
  4. 2024Chain-of-Thought Unfaithfulness as Disguised Accuracy
  5. 2026Beyond the Sycophancy Score: How Task, Model, and Pressure Shape LLM Yielding
  6. 2026Do Models Know Why They Changed Their Mind? Interpretability and Faithfulness of Chain-of-Thought Under Knowledge Conflict
  7. 2025Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps
  8. 2025Chain-of-Thought Reasoning In The Wild Is Not Always Faithful