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

The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think

ICLR 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 Encyclopedia, a bottom-up LLM-driven pipeline that extracts, clusters, and rubricizes reasoning-strategy criteria from long CoTs, then uses the resulting strategy predictions to steer models toward better accuracy and safety and studies how training data format shapes reasoning.

Problem

Existing analyses of long chain-of-thought reasoning rely on predefined strategy categories (e.g., backtracking, verification) that are constrained by human intuition and miss the diversity of model behaviors, and it is unclear how strategies can be controlled to improve performance.

Approach

Replace fixed taxonomies with a bottom-up pipeline: LLM-brainstormed contrastive criteria are embedded, hierarchically clustered via medoids, turned into binary rubrics, and used both to profile responses and to predict/steer which strategy yields correct or safe answers per question.

Key results

• Guiding models with question-specific optimal reasoning strategies raises GPQA-Diamond accuracy from 72.7% to 79.1%, and XSTest/WildGuard safety from 91.1% to 94.0% and 89.1% to 92.9%.
• Human annotators judged the bottom-up criteria as reasonable in 92-97% of cases versus 51% for top-down cognitive behavior labels.
• Training data format has much larger effects on reasoning strategies than domain (Cohen's d up to 1.5 for format vs. below 0.2 for domain); linear weight interpolation between MC and FF models yields smoothly varying strategies.

Surprising findings

• Cognitive behavior frequencies (verification, backtracking, etc.) barely differ across reasoning models, suggesting the conventional taxonomy has limited discriminative power.
• Unoptimal strategy instructions sometimes perform worse than giving no instruction at all.

Limitations

• Classification relies on GPT-4o as evaluator, which may carry its biases and not fully reflect human judgment.
• Experiments limited to three benchmarks and three model families; generality beyond these is untested.
• Strategy-guided improvements depend on models reliably following stylistic instructions, limiting applicability to weaker or non-instruction-tuned models.

Experimental setup

Datasets: GPQA-Diamond MMLU-Redux MATH-500 XSTest WildGuard Arena-Hard NuminaMath SciQ QASC OpenBookQA CommonsenseQA ARC-Challenge

Benchmarks: GPQA-Diamond MMLU-Redux MATH-500 XSTest WildGuard Arena-Hard

Metrics: accuracy safe response ratio Cohen's d chi-squared p-value human-judged plausibility ratio

Baselines: cognitive behavior classification (top-down predefined behaviors) zero-shot prompting classifier few-shot prompting classifier

Builds on: chain-of-thought prompting GRPO cognitive behaviors in LLM reasoning (verification, backtracking, subgoal settin hierarchical agglomerative clustering

Closest papers by meaning

• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.93)
• To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning (ICLR 2025 · similarity 0.92)
• Understanding Chain-of-Thought in LLMs through Information Theory (ICML 2025 · similarity 0.92)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.92)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.91)
• CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs (ICLR 2026 · similarity 0.91)