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

CoT-Evo: Evolutionary Distillation of Chain-of-Thought for Scientific Reasoning

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.

Proposes CoT-Evo, an evolutionary CoT distillation framework that evolves a pool of multi-teacher reasoning trajectories via novelty-driven selection, reflective recombination and mutation, and fine-tunes compact models on the evolved data for scientific reasoning.

Problem

Distilling chain-of-thought from fallible LLM teachers into small models yields low-quality, superficial or incorrect reasoning data in scientific domains, limiting student performance.

Approach

Treats CoT distillation as an evolutionary process: a diverse multi-thinker candidate pool, knowledge augmentation, a composite fitness (exact match, length, knowledge-use judge), novelty-plus-local-competition Pareto selection, and LLM-driven recombination and mutation that fuse steps across chains into a single improved chain (intra-chain aggregation).

Key results

• On BioProBench, CoT-Evo yields relative gains of 12.6% over single-teacher and 8.4% over multi-teacher distillation; on ChemCoTBench gains reach 27.0% and 19.3%.
• Qwen3-8B fine-tuned on CoT-Evo data reaches 0.649 PQA accuracy and 0.544 ORD EM on BioProBench, versus 0.601 and 0.368 for its single-teacher baseline.
• CoT-Evo data keeps over 70% usable (correct) data (0.729 on BioProBench, 0.704 on ChemCoTDataset) versus at most 0.536 for multi-teacher baseline.

Surprising findings

• Best-of-K sampling with 21 samples per teacher gives only modest gains, suggesting improvements come from evolution rather than sampling budget.

Limitations

• Multiple LLM thinkers and iterative evaluation incur substantial computational and time cost.
• LLM-based reward signals (knowledge judge, GPT-5 quality evaluation) may carry biases.
• Evaluated only on two scientific benchmarks (chemistry and biology protocols).

Experimental setup

Datasets: ChemCoTDataset BioProBench Llama-Nemotron-Science

Benchmarks: BioProBench ChemCoTBench

Metrics: Accuracy Exact Match F1 MAE Tanimoto molecule similarity success rate FTS fingerprint similarity pass rate quality score win rate

Baselines: Single Teacher (ST) Multi Teacher (MT) Best-of-K (BoK) DeepSeek-R1 Qwen3-235B-A22B-think Qwen3-32B-think Llama4-Scout (backward, CoT, ToT)

Builds on: Novelty Search with Local Competition DeepSeek-R1 Chain-of-Thought Tree-of-Thought Llama-Nemotron-Science

Closest papers by meaning

• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.90)
• Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillation (NeurIPS 2026 · similarity 0.90)
• Mind the Gap: Bridging Thought Leap for Improved Chain-of-Thought Tuning (NeurIPS 2025 · similarity 0.89)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.89)
• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.89)
• Rethinking Chain-of-Thought from the Perspective of Self-Training (ICML 2025 · similarity 0.89)