AI Paper Atlas

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

Hidden Patterns in Chain-of-Thought Reasoning

Gregor Bachmann, Yichen Jiang, Seyed-Mohsen Moosavi-Dezfooli, Moin Nabi

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis

Read the paperCheck your idea against 137,217 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.

Analyzes chain-of-thought traces to uncover what drives CoT success in LLMs; the abstract is truncated so details beyond the opening motivation are unknown.

Problem

The driving forces behind the success of chain-of-thought prompting in LLMs remain largely unclear.

Approach

An in-depth analysis of CoT traces to identify the underlying factors behind CoT reasoning success; the specific method is not stated in the available text.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Examine per-chunk success probability (potential) of CoT traces on AIME-2024non-reasoning and reasoning LLMs · AIME-2024fraction of monotonic potential curves: 15% to 45% monotonic across modelsMixed
Flag late-spike (guess) samples when computing pass@kQwen2.5-1.5B and 7B · AIME-2025pass@kHurt
Greedily select the highest-potential candidate chunk at each step (optimized CoT)Qwen2.5-1.5B · AIME-2025potential monotonicityImproved
Provide 20% of a stronger model's partial CoT to Qwen3-0.6BQwen3-0.6B · AIME-2025accuracyImproved
Provide partial CoT from GPT-OSS-20B instead of same-family Qwen3-32BQwen3-0.6B · AIME-2025accuracyMixed
Provide partial summary CoT from Qwen3-235B to non-reasoning modelsQwen2.5-7B and Qwen2.5-72B · AIME-2024accuracyImproved
Use reasoning models (Qwen3) rather than non-reasoning modelsAIME-2024 CoT traces · AIME-2024% CoTs with tangent: Qwen3-0.6B 41% vs Qwen2.5-1.5B 5%Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy