In a nutshell: Evolution Fine-Tuning teaches language models to generalize solution strategies across different optimization problems while achieving 10.22 percent better results than baseline models.
Researchers have developed a method that trains language models to generalize solution strategies from optimization tasks rather than starting each problem from scratch. The model learns which parts of a solution to modify and when to take steps backward.
The new Evolution Fine-Tuning (EFT) method addresses a previously unsolved problem: while Large Language Models (LLMs) already achieve top results in evolutionary search algorithms for GPU kernel design, mathematical problems, and combinatorial puzzles, the knowledge is discarded after each individual task. Every new problem is tackled from the ground up, even though the strategies acquired during the search could be transferable to other problems.
In EFT, the evolution paths from optimization tasks are used directly as training signals. The researchers constructed the Finch Collection, a dataset containing 156,000 evolution trajectories spanning ten domains and 371 optimization tasks. With this material, open-source models ranging from 2 to 9 billion parameters were fine-tuned. The method teaches the model itself which mutations are sensible and when a step backward is necessary to avoid local minima — rather than hardcoding this logic only into the external search structure.
In evaluation on 22 unseen tasks, an average improvement of 10.22 percent compared to baseline models was demonstrated. Combined with test-time reinforcement learning, the model achieved state-of-the-art results on two circle packing problems and outperformed the baseline on the Erdős Minimum Overlap Problem. EFT thus functions as a “training phase” for universal discovery agents that do not approach new problems without prior knowledge.
Source: arxiv.org · Published June 26, 2026
Lumi AI News — AI-assisted curation according to Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.2.