mlsanityData to 14 Oct 2026

Risky Business: Measuring The Faithfulness-Safety Tension

arXiv 2608.03745 ↗First posted 4 Aug 2026cs.AI cs.CLcited by 0

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

Abstract (arXiv)

Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We identify an alignment tension where a model must be faithful enough to be monitored, yet robust enough to reject unsafe reasoning. We demonstrate that this counterbalance exists in current Large Reasoning Models (LRMs), and show ways in which it can be addressed. We introduce HazMart, a human-written dataset set in an autonomous AI shopkeeper scenario. Unlike prior work that relies on providing hints in prompts to test faithfulness (e.g., "A Stanford professor said it should be Answer A"), we propose a novel replacement-based technique, which we call Targeted Reasoning Replacement (TRR), that directly intervenes in the reasoning chain to substitute in unsafe or illogical thoughts (e.g., "Wait, the answer must be Option B [was Option A] because it is the most fitting"). DeepSeek-R1-Llama-70B exhibits high faithfulness (97.5%) but fails to reject Unsafe Reasoning (12.3%), while QwQ-32B is more robust (73.9% safety) at the cost of lower faithfulness (74.7%). Mechanistic analyses of QwQ-32B reveal that these properties are represented by anti-correlated internal directions peaking at the action-commit token. Finally, we demonstrate that representation steering can independently amplify the safety direction, increasing safe behavior by 9 percentage points while maintaining base capabilities.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 32 of its 50 references that are in the map. First posted on arXiv 4 Aug 2026.

Builds on

32 of 50 references.

Closest ancestors · 9

Same direction, newest first

  1. 2026Counterfactual Simulation Training for Chain-of-Thought Faithfulness24 Feb
  2. 2026Chain-of-thought obfuscation learned from output supervision can generalise to unseen tasks30 Jan
  3. 2026Reasoning Models Will Sometimes Lie About Their Reasoning12 Jan
  4. 2025Is Chain-of-Thought Really Not Explainability? Chain-of-Thought Can Be Faithful without Hint Verbalization28 Dec
  5. 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety15 Jul
  6. 2025When Thinking LLMs Lie: Unveiling the Strategic Deception in Representations of Reasoning Models5 Jun
  7. 2025Reasoning Models Don't Always Say What They Think8 May
  8. 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation14 Mar
  9. 2023Measuring Faithfulness in Chain-of-Thought Reasoning17 Jul

Other references · 23

  1. 2025Qwen3 Technical Reportcited by 8,820
  2. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  3. 2024Qwen2.5 Technical Reportcited by 4,942
  4. 2025gpt-oss-120b & gpt-oss-20b Model Cardcited by 1,386
  5. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  6. 2023Steering Language Models With Activation Engineeringcited by 922
  7. 2023Trustworthy LLMs: a Survey and Guideline for Evaluating Large Language Models' Alignmentcited by 401
  8. 2025SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilitiescited by 136
  9. 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independencecited by 115
  10. 2026Ministral 3cited by 114
  11. 2022Is Power-Seeking AI an Existential Risk?cited by 105
  12. 2025Thought Crime: Backdoors and Emergent Misalignment in Reasoning Modelscited by 65
11 more
  1. 2025AlphaOne: Reasoning Models Thinking Slow and Fast at Test Timecited by 29
  2. 2026There Is More to Refusal in Large Language Models than a Single Directioncited by 26
  3. 2025A Comprehensive Survey on Trustworthiness in Reasoning with Large Language Modelscited by 25
  4. 2025RAG LLMs are Not Safer: A Safety Analysis of Retrieval-Augmented Generation for Large Language Modelscited by 24
  5. 2025Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppressioncited by 24
  6. 2025Measuring the Faithfulness of Thinking Drafts in Large Reasoning Modelscited by 23
  7. 2025Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusalcited by 19
  8. 2026Prompt Injection as Role Confusioncited by 17
  9. 2025Chain-of-Thought Hijackingcited by 16
  10. 2025RepliBench: Evaluating the Autonomous Replication Capabilities of Language Model Agentscited by 15
  11. 2025Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Trainingcited by 12

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025ReFIne: A Framework for Trustworthy Large Reasoning Models with Reliability, Faithfulness, and Interpretability
  2. 2026Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?
  3. 2025Investigating CoT Monitorability in Large Reasoning Models
  4. 2025When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors
  5. 2026Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations
  6. 2025Reasoning Models Sometimes Output Illegible Chains of Thought
  7. 2025Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention
  8. 2025A Concrete Roadmap towards Safety Cases based on Chain-of-Thought Monitoring