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

Researchers have introduced Evolution Fine-Tuning (EFT), a mid-training paradigm that teaches large language models to iteratively improve solutions across disparate optimization tasks rather than starting from scratch each time, according to a preprint posted to arXiv [1]. The approach addresses a structural limitation in earlier work, where LLMs were integrated into evolutionary search scaffolds that tackled one target task at a time. Once a run finished, the accumulated search experience was discarded, leaving the capability to decide which part of a solution to mutate or when to backtrack entirely in the scaffold rather than in the model itself [1]. EFT converts evolutionary search trajectories into supervised training data so the model internalizes the process of evolving a solution [1]. To train the models, the team constructed the Finch Collection, a dataset of 156,000 trajectories spanning 10 domains and 371 optimization tasks [1]. The domains include mathematical discovery, competitive programming, heuristic optimization, numerical algorithm optimization, symbolic regression, GPU kernel optimization, constructive search, and biological denoising [3]. Trajectories were collected using the OpenEvolve scaffold with Qwen3.5-397B-A17B as the teacher mutation operator, running under two strategies — diff-based edit for exploitation and full rewrite for exploration — at temperature 0.7 and top-p 0.95 [5]. After filtering to remove systematic errors and unrecoverable cases, 156,731 transitions were retained, each labeled by its score delta [5]. Open-source LLMs ranging from 2 billion to 9 billion parameters were fine-tuned on the dataset [1]. On 22 held-out tasks, the EFT models outperformed their base counterparts by an average of 10.22 percent [1]. When combined with test-time reinforcement learning, the fine-tuned model matched state-of-the-art performance on two circle-packing tasks and surpassed its base-model counterpart on the Erdős minimum-overlap problem [1]. The researchers frame EFT as a “practice phase” for general-purpose discovery agents that do not solve new problems from scratch [1]. The dataset is intentionally heterogeneous and highly imbalanced across task groups: symbolic regression tasks contribute the most trajectories, while competitive programming contributes the largest share of individual tasks at 172 from the FrontierCS benchmark [5]. Two teacher variants of the Finch Collection have also been released, one using GPT-5.4 with 5,535 trajectories across 14 tasks in three groups, and another using Gemini-3-Flash with 8,120 trajectories across 27 tasks in five groups [9][10]. The Gemini-3-Flash variant shows a different improvement profile, with 53.7 percent of transitions labeled as regressed and 43.1 percent as improved, compared with the main collection’s 39.4 percent improved and 41.3 percent regressed [5][10].