Research map / Large language models / Instruction and fine-tuning

Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks

NeurIPS 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data · 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 Evolution Fine-Tuning (EFT), a mid-training method that fine-tunes Qwen3.5 models (2B-9B) on 156K OpenEvolve search trajectories across 371 optimization tasks to internalize cross-task evolutionary discovery ability.

Problem

LLM-driven evolutionary search for optimization tasks (math conjectures, GPU kernels, algorithm engineering) relies on external scaffolds and frontier models; the iterative discovery skill is not internalized into smaller open models and does not transfer across tasks.

Approach

Treat evolutionary search trajectories (parent, history, feedback -> child) as SFT supervision so the model learns to act as a stronger mutation operator across tasks; optionally add KTO on improved vs regressed trajectories and test-time RL as mid-training.

Key results

• EFT models surpass same-size base models by 10.22% on average across 22 held-out tasks (Table reports +10.24% for 9B; +1.56% for 2B, +3.40% for 4B, +3.17% for 8B).
• Scaling training tasks from 15 to 355 improves held-out performance by 14.1% on average.
• With test-time RL, EFT-9B matches state-of-the-art on two circle-packing tasks and improves Erdos by 3.2% over base.

Surprising findings

• Qwen3-8B baseline on Erdos (0.4036) is outperformed by the 4B EFT model (0.3865), i.e., a smaller EFT model matches a larger non-EFT model.
• Adding KTO on regressed trajectories after SFT further improves discovery; the 8B KTO model surpasses the best human score on AC1 and AC2.

Limitations

• Trained and evaluated only with OpenEvolve scaffold; may not generalize to stronger scaffolds like EvoX.
• Test-time RL experiments only on mathematical tasks; smaller models still lag TTT-Discover with GPT-OSS-120B.
• Single-turn formulation; no multimodal or multi-turn lineage modeling.

Experimental setup

Datasets: OpenEvolve trajectories from AlphaEvolve math problems FrontierCS ALE-Bench AlgoTune GPU Mode LLM-SRBench OpenEvolve function minimization scRNA-seq denoising Erdos variants

Benchmarks: Erdos minimum overlap First autocorrelation inequality Second autocorrelation inequality Circle packing n=26 Hadamard ahc039 ahc058 EPLB PRISM LLM-SQL Transaction FrontierCS P263/P301-305

Metrics: task-specific score (max within T=100 iterations) relative improvement over base (Avg. Gain)

Baselines: OpenEvolve with base Qwen3.5 models Qwen3-8B Claude-Opus-4.6 Gemini-3-Pro GPT-5 TTT-Discover

Builds on: OpenEvolve TTT-Discover KTO LLaMA-Factory AlphaEvolve

Closest papers by meaning

• EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling (NeurIPS 2026 · similarity 0.89)
• ESSAM: A Novel Competitive Evolution Strategies Approach to Reinforcement Learning for Memory Efficient LLMs Fine-Tuning (NeurIPS 2026 · similarity 0.89)
• MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.87)
• ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution (NeurIPS 2024 · similarity 0.87)
• SolverLLM: Leveraging Test-Time Scaling for Optimization Problem via LLM-Guided Search (NeurIPS 2025 · similarity 0.87)
• ThetaEvolve: Test-time Learning on Open Problems (ICML 2026 · similarity 0.87)