Hidden Reasoning Must Leak, but Need Not Be Readable: Fundamental Opportunities and Limits for Chain-of-Thought Monitoring
On the map: LLM reasoning → Prompting and chain-of-thought reasoning
Abstract (arXiv)
Can reasoning models trick chain of thought (CoT) monitors and perform hidden computation without revealing it in their thinking traces? We show that the answer depends on the underlying task difficulty and the model size. Simple computations can be performed covertly; however, beyond a threshold depending on model size, successfully solving the task necessarily leaks a near-linear amount of information about the covert task input into the CoT. Therefore, sufficiently complex hidden computation always leaves an information-theoretic footprint. However, concerningly, this leakage need not be readable: Under plausible cryptographic assumptions, even a one-layer Transformer can encrypt its reasoning online so that no polynomial-time monitor can extract information about the hidden computation. Overall, our theoretical and empirical results provide a holistic view of both the opportunities and the limitations of CoT monitoring.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 24 of its 79 references that are in the map. First posted on arXiv 29 Sep 2026.
Builds on
24 of 79 references.
Closest ancestors · 9
Same direction, newest first
- 2026Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?31 Mar
- 2026Quantifying the Necessity of Chain of Thought through Opaque Serial Depth10 Mar
- 2026Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory20 Feb
- 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety15 Jul
- 2025When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors7 Jul
- 2025Exact Expressive Power of Transformers with Padding25 May
- 2025Reasoning Models Don't Always Say What They Think8 May
- 2023Preventing Language Models From Hiding Their Reasoning27 Oct
- 2023Measuring Faithfulness in Chain-of-Thought Reasoning17 Jul
Other references · 15
- 2020On the Computational Power of Transformers and its Implications in Sequence Modelingcited by 85
- 2024Transformers Can Do Arithmetic with the Right Embeddingscited by 77
- 2025SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agentscited by 58
- 2023Masked Hard-Attention Transformers Recognize Exactly the Star-Free Languagescited by 47
- 2025Evaluating Frontier Models for Stealth and Situational Awarenesscited by 30
- 2023Average-Hard Attention Transformers are Constant-Depth Uniform Threshold Circuitscited by 26
- 2024Lessons from Studying Two-Hop Latent Reasoningcited by 15
- 2024Excuse me, sir? Your language model is leaking (information)cited by 11
- 2026Tool Use Enables Undetectable Steganography in Multi-Agent LLM Systemscited by 5
- 2026Not All LLM Reasoning is Visible in the Chain-of-Thoughtcited by 5
- 2025Robust Steganography from Large Language Modelscited by 4
- 2025A completely uniform transformer for paritycited by 3
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback
- 2026Selective Disclosure of Hidden Directives in Reasoning Models: Behavioral Asymmetry and Steering
- 2026Does Out-of-Sight Equal Out-of-Mind in CoT Monitorability?
- 2026NEST: Nascent Encoded Steganographic Thoughts
- 2026Training on Documents About Monitoring Leads to CoT Obfuscation
- 2025CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring
- 2026Monitor Jailbreaking: Evading Chain-of-Thought Monitoring Without Encoded Reasoning
- 2026Chain-of-thought obfuscation learned from output supervision can generalise to unseen tasks