mlsanityData to 14 Oct 2026

Reasoning Models Struggle to Control their Chains of Thought

arXiv 2603.05706 ↗First posted 5 Mar 2026cs.AIcited by 25

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

Abstract (arXiv)

Chain-of-thought (CoT) monitoring is a promising tool for detecting misbehaviors and understanding the motivations of modern reasoning models. However, if models can control what they verbalize in their CoT, it could undermine CoT monitorability. To measure this undesirable capability -- CoT controllability -- we introduce the CoT-Control evaluation suite, which includes tasks that require models to solve problems while adhering to CoT instructions, e.g., reasoning about a genetics question without using the word 'chromosome'. We show that reasoning models possess significantly lower CoT controllability than output controllability; for instance, Claude Sonnet 4.5 can control its CoT only 2.7% of the time but 61.9% when controlling its final output. We also find that CoT controllability is higher for larger models and decreases with more RL training, test-time compute, and increased problem difficulty. CoT controllability failures extend even to situations in which models are given incentives (as opposed to direct requests) to evade CoT monitors, although models exhibit slightly higher controllability when they are told they are being monitored. Similarly, eliciting controllability by adversarially optimizing prompts does not meaningfully increase controllability. Our results leave us cautiously optimistic that CoT controllability is currently unlikely to be a failure mode of CoT monitorability. However, the mechanism behind low controllability is not well understood. Given its importance for maintaining CoT monitorability, we recommend that frontier labs track CoT controllability in future models.

Key facts. Cited by 25 papers 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 39 of its 49 references that are in the map. First posted on arXiv 5 Mar 2026.

Builds on

39 of 49 references.

Closest ancestors · 11

Same direction, newest first

  1. 2025Monitoring Monitorability20 Dec
  2. 2025Can Reasoning Models Obfuscate Reasoning? Stress-Testing Chain-of-Thought Monitorability21 Oct
  3. 2025Output Supervision Can Obfuscate the Chain of Thought11 Oct
  4. 2025LLMs Can Covertly Sandbag on Capability Evaluations Against Chain-of-Thought Monitoring31 Jul
  5. 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety15 Jul
  6. 2025When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors7 Jul
  7. 2025Large language models can learn and generalize steganographic chain-of-thought under process supervision2 Jun
  8. 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation14 Mar
  9. 2022Large Language Models are Zero-Shot Reasoners24 May
  10. 2022Self-Consistency Improves Chain of Thought Reasoning in Language Models21 Mar
  11. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Models28 Jan

Other references · 28

  1. 2025Qwen3 Technical Reportcited by 8,820
  2. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  3. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  4. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  5. 2023Tree of Thoughts: Deliberate Problem Solving with Large Language Modelscited by 3,211
  6. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  7. 2022Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language modelscited by 1,998
  8. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,986
  9. 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,848
  10. 2025gpt-oss-120b & gpt-oss-20b Model Cardcited by 1,386
  11. 2024OpenAI o1 System Cardcited by 1,244
  12. 2022Red Teaming Language Models with Language Modelscited by 1,026
16 more
  1. 2024Large Language Monkeys: Scaling Inference Compute with Repeated Samplingcited by 991
  2. 2025GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learningcited by 542
  3. 2025Humanity's Last Examcited by 491
  4. 2025OpenThoughts: Data Recipes for Reasoning Modelscited by 304
  5. 2025Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMscited by 289
  6. 2024Teaching Large Language Models to Reason with Reinforcement Learningcited by 160
  7. 2023Taken out of context: On measuring situational awareness in LLMscited by 143
  8. 2025Olmo 3cited by 129
  9. 2025Large Language Models Often Know When They Are Being Evaluatedcited by 128
  10. 2025Stress Testing Deliberative Alignment for Anti-Scheming Trainingcited by 86
  11. 2025Magistralcited by 60
  12. 2023Characterizing Manipulation from AI Systemscited by 43
  13. 2025Monitoring Decomposition Attacks in LLMs with Lightweight Sequential Monitorscited by 25
  14. 2025ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoningcited by 13
  15. 2025SAGE-Eval: Evaluating LLMs for Systematic Generalizations of Safety Factscited by 8
  16. 2024Emotional Manipulation Through Prompt Engineering Amplifies Disinformation Generation in AI Large Language Modelscited by 5

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 25.

25 papers.

Sort:

By area:

  1. 2026Automated Researchers Can Mitigate Well-characterized Alignment Failurescited by 5
  2. 2026"Did you lie?" Evaluating Lie Detectors across Model Scale and Belief-Verified Model Organismscited by 4
  3. 2026Training on Documents About Monitoring Leads to CoT Obfuscationcited by 3
  4. 2026Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoningcited by 3
  5. 2026Knowledge-Based Zero-Replay Debugging of Multi-Agent LLM Tracescited by 2
  6. 2026Chain-of-Thought Monitoring Can Be Unreliable in Implicit-Influence Settingscited by 2
  7. 2026Bootstrapped Monitoring: Leveraging Transparent Reasoning to Oversee Stronger AI Agentscited by 1
  8. 2026Length Penalties Make Chain-of-Thought Less Monitorablecited by 1
  9. 2026Measuring Activation Control in Large Language Modelscited by 1
  10. 2026AISafetyBenchExplorer: A Metric-Aware Catalogue of AI Safety Benchmarks Reveals Fragmented Measurement and Weak Benchmark Governancecited by 0
  11. 2026SCOReD: Student-Aware CoT Optimization for Recommendation Distillationcited by 0
  12. 2026INTENT-AS-A-TOOL Makes it Easy to Track Agentic Misalignmentcited by 0
13 more
  1. 2026In-Context Neurofeedback: Can LLMs Control Their Internal Representations through Privileged Access?cited by 0
  2. 2026Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrievalcited by 0
  3. 2026Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injectioncited by 0
  4. 2026When the Agent Becomes the Kernel: A Systematization of Security on the Path to AI-Native Operating Systemscited by 0
  5. 2026Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcementcited by 0
  6. 2026OPBackdoor: Opportunistic Backdoors via Alibi-Aligned Reasoningcited by 0
  7. 2026Control-Token Injection Suppresses Chain-of-Thought and Defeats Reasoning-Based Oversight in Tool-Using Agentscited by 0
  8. 2026When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitmentcited by 0
  9. 2026Towards Mitigating Deceptive Safety Alignment in Large Reasoning Modelscited by 0
  10. 2026What Was Said, Not What Was 'Thought': Type-6 Logic for CoT Verificationcited by 0
  11. 2026Learning Steganography Is Easy, Learning Steganographic Reasoning Is Hardcited by 0
  12. 2026Monitorability Disposition in Large Reasoning Modelscited by 0
  13. 2026AgentTime: Can Agents Estimate and Control Their Own Runtime?cited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Investigating CoT Monitorability in Large Reasoning Models
  2. 2026Does Out-of-Sight Equal Out-of-Mind in CoT Monitorability?
  3. 2025Reasoning Models Don't Always Say What They Think
  4. 2026Chain of Thought Monitoring Can Fail Even When CoT Is Necessary
  5. 2025The Ends Justify the Thoughts: RL-Induced Motivated Reasoning in LLM CoTs
  6. 2026Diagnosing Pathological Chain-of-Thought in Reasoning Models
  7. 2026Evading Chain-of-Thought Monitoring Through Model Poisoning
  8. 2025Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity