Research map / Large language models / Reasoning and chain of thought

Reasoning Models Struggle to Control their Chains of Thought

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces CoT-Control, a 14,076-instance suite measuring whether reasoning models can control their chain-of-thought under instructions, finding CoT controllability is far lower than output controllability and decreases with RL training, reasoning length, and difficulty.

Problem

Chain-of-thought monitoring relies on reasoning traces being faithful and hard to manipulate; if models can control what they verbalize in their CoT, they could evade monitors. The paper asks how well current reasoning models can control their CoTs.

Approach

Measure CoT controllability as a proxy by instructing models to follow CoT-level constraints (suppress keywords, add content, stylistic changes) while solving real benchmark questions, scored programmatically, and compare against output controllability and external-reasoning baselines; then study how factors like RL training, reasoning effort, length, difficulty, and situational awareness affect i

Key results

• Claude Sonnet 4.5 controls its CoT only 2.7% of the time versus 61.9% when controlling its final output.
• DeepSeek R1 CoT controllability can be as low as 0.1%, over two orders of magnitude below its output controllability.
• OLMo-3-7B-RL-Zero-Math CoT controllability decreases by over an order of magnitude across RL checkpoints; GPT-OSS-120B reaches 5.9% CoT controllability vs 0.9% for 20B.

Surprising findings

• Stronger models often mention the CoT constraint in their reasoning (meta-discussion), which a monitor could easily flag; for ignore-question instructions meta-discussion is ≥90%.
• Automated GEPA-style adversarial pressure prompts do not reliably transfer to unseen controllability types, with only 3 of 8 models improving, each by under 5 pp.

Limitations

• Controllability is measured via instruction-following proxies rather than spontaneous misaligned evasion of monitors.
• Elicitation limited to prompting; RL or SFT-based elicitation not explored.
• Breadth of control types is limited.

Experimental setup

Datasets: GPQA MMLU-Pro Humanity's Last Exam BFCL SWE-Bench Verified

Benchmarks: CoT-Control GPQA MMLU-Pro Humanity's Last Exam BFCL SWE-Bench Verified

Metrics: controllability score meta discussion rate controllability score without meta-discussion pass@10 accuracy

Baselines: output controllability external-CoT controllability GPT-4o Llama 3.3 70B Instruct DeepSeek Chat v3

Builds on: GEPA RLVR OLMo-3

Closest papers by meaning

• Reasoning Models Sometimes Output Illegible Chains of Thought (NeurIPS 2025 · similarity 0.90)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.89)
• Monitoring Monitorability (ICML 2026 · similarity 0.89)
• CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring (NeurIPS 2025 · similarity 0.88)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.88)
• To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning (ICLR 2025 · similarity 0.88)