CoT Monitoring

CoT: Chain-of-Thought

Chain-of-thought monitoring asks whether a model’s visible reasoning reveals the factors behind its answers and actions, so people or automated monitors can detect errors, unsafe behavior, and deception. Researchers test this with benchmarks and interventions such as hint changes, counterfactuals, and adversarial prompts, and build monitors using reasoning text, model activations, gradients, or action signals. Training approaches try to make reasoning more faithful or legible through targeted rewards, behavior cues, and monitor-focused fine-tuning. A central tension is that traces can omit influential information or be deliberately shaped to evade oversight, with monitorability varying by task, language, and training method.

Latest papers

40 papers11 papers in the last four weeks, up 175% on the four weeks before. 0.1% of all new papers.

Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?
Mar 31, 2026Max Kaufmann, David Lindner, Roland S. Zimmermann +1CoT Monitoring CoT Supervision
Google DeepMind
MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models
Mar 30, 2026Han Wang, Yifan Sun, Brian Ko +8CoT Faithfulness LLM Evaluation
University of Illinois Urbana-Champaign · University of Washington · University of California San Diego
Beyond Content Safety: Real-Time Monitoring for Reasoning Vulnerabilities in Large Language Models
Mar 26, 2026Xunguang Wang, Yuguang Zhou, Qingyue Wang +5LLM Safety Monitoring LLM Safety
The Hong Kong University of Science and Technology, Hong Kong, China · Zhejiang University of Technology, Hangzhou, Zhejiang Province, China
DeceptGuard :A Constitutional Oversight Framework For Detecting Deception in LLM Agents
Mar 14, 2026Snehasis MukhopadhyayDeception in Language Models AI Agent Monitoring
Indian Institute of Information Technology Kalyani
Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback
Mar 14, 2026Thomas Jiralerspong, Flemming Kondrup, Yoshua BengioCoT Monitoring AI Agent Monitoring
Mila · Université de Montréal · McGill University +1
Reasoning Models Struggle to Control their Chains of Thought
Mar 5, 2026Chen Yueh-Han, Robert McCarthy, Bruce W. Lee +5CoT Monitoring Large Reasoning Models
NYU, MATS · UCL, MATS · UPenn, MATS +2
Counterfactual Simulation Training for Chain-of-Thought Faithfulness
Feb 24, 2026Peter Hase, Christopher PottsCoT Faithfulness CoT Monitoring
Stanford University
Analyzing and Improving Chain-of-Thought Monitorability Through Information Theory
Feb 20, 2026Usman Anwar, Tim Bakker, Dana Kianfar +2CoT Monitoring
University of Cambridge, work done while at Qualcomm AI Research. · Qualcomm AI Research.
NEST: Nascent Encoded Steganographic Thoughts
Feb 15, 2026Artem KarpovLanguage Model Safety Evaluation LLM Auditing
Independent.
The Ends Justify the Thoughts: RL-Induced Motivated Reasoning Fools CoT Monitors
Oct 20, 2025Nikolaus Howe, Micah CarrollLLM Safety Monitoring CoT Monitoring
Mila, Université de Montréal · UC Berkeley

Mar 31, 2026J/K move · Enter open · S save

Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?

Max Kaufmann, David Lindner, Roland S. Zimmermann, and Rohin Shah

Google DeepMind

CoT Monitoring CoT Supervision Language Model Post-Training LLM Auditing RL for Language Model Reasoning

Open paper AskPDF arXiv

Chain-of-Thought (CoT) monitoring, in which automated systems monitor the CoT of an LLM, is a promising approach for effectively overseeing AI systems. However, the extent to which a model's CoT helps us oversee the model - the monitorability of the CoT - can be affected by training, for instance by the model learning to hide important features of its reasoning. We propose and empirically validate a conceptual framework for predicting when and why this occurs. We model LLM post-training as an RL environment where the reward decomposes into two terms: one term depending on final outputs and another term depending on the CoT. Our framework allows us to classify these two terms as "aligned", "orthogonal", or "in-conflict" before training. We predict that training with in-conflict terms will reduce monitorability, orthogonal terms will not affect it, and aligned terms will improve it. To validate our framework, we use it to classify a set of RL environments, train LLMs within those environments, and evaluate how training affects CoT monitorability. We find that (1) training with "in-conflict" reward terms reduces CoT monitorability and (2) optimizing in-conflict reward terms is difficult.