Hidden Reasoning Must Leak, but Need Not Be Readable: Fundamental Opportunities and Limits for Chain-of-Thought Monitoring

arXiv:2609.37312 · cs.LG, cs.AI, cs.CL · Submitted 2026-09-29 · Read on arXiv

cs.LG, cs.AI, cs.CL

Submitted: 2026-09-29

Updated: 2026-09-29

Code: https://github.com/ggml-org/llama.cpp

Terminology

• Accuracy — 161× in this paper · explained in 1 episode
• Transformer — 146× in this paper · explained in 2 episodes
• Filler Tokens — 92× in this paper · explained in 2 episodes
• Trace — 23× in this paper · explained in 2 episodes
• Ablation — 22× in this paper · explained in 1 episode
• Embeddings — 6× in this paper · explained in 2 episodes
• Generalization — 4× in this paper · explained in 13 episodes
• Decode — 3× in this paper · explained in 2 episodes
• Calibration — 2× in this paper · explained in 7 episodes
• Multi-hop Reasoning — 2× in this paper · explained in 3 episodes
• Transformer Architecture — 2× in this paper · explained in 2 episodes
• AI Agents — 1× in this paper · explained in 2 episodes
• Chain-of-Thought Prompting — 1× in this paper · explained in 2 episodes
• Conference on Computational Complexity — 1× in this paper
• Consistency — 1× in this paper · explained in 3 episodes
• Hypergraph — 1× in this paper · explained in 2 episodes
• Latent Reasoning — 1× in this paper · explained in 2 episodes
• Quantization — 1× in this paper · explained in 12 episodes

Sources

• Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
• Not All LLM Reasoning is Visible in the Chain-of-Thought
• Lessons from Studying Two-Hop Latent Reasoning
• Reading Between the Dots: Decoding Hidden Computation across Filler Tokens
• Quantifying the Necessity of Chain of Thought through Opaque Serial Depth
• Reasoning Models Don't Always Say What They Think
• Exploring the Cryptographic Limits of Transformer Networks
• When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors
• Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?
• Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
• A completely uniform transformer for parity
• SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
• Measuring Faithfulness in Chain-of-Thought Reasoning
• ResearchArena: Evaluating Sabotage and Monitoring in Automated AI R&D
• Transformers Can Do Arithmetic with the Right Embeddings
• Robust Steganography from Large Language Models
• Evaluating Frontier Models for Stealth and Situational Awareness
• Tool Use Enables Undetectable Steganography in Multi-Agent LLM Systems
• Preventing Language Models From Hiding Their Reasoning
• Average-Hard Attention Transformers are Constant-Depth Uniform Threshold Circuits

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks