Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks
TLDR
Introduces Evolution Fine-Tuning (EFT) to teach LLMs to evolve solutions across optimization tasks, achieving cross-task generalization.
Reasoning
The paper presents a novel mid-training paradigm that leverages evolutionary search trajectories to enable cross-task generalization in LLMs, supported by a large dataset and empirical improvements. However, the focus is on optimization tasks rather than broader scientific discovery or research automation, limiting direct relevance to some keywords.
Read-first score
Read-first score 41.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.
Field roles
Rank sensitivity
Stability: volatile; rank range: 35.
Keyword Scores
Deep Analysis
Innovations
- Evolution Fine-Tuning (EFT): a mid-training paradigm that teaches LLMs to evolve solutions across tasks by converting evolutionary search trajectories into supervision.
- Finch Collection: a dataset of 156K evolutionary search trajectories spanning 10 domains and 371 optimization tasks.
- Transferring the capability of iterative solution evolution from the search scaffold to the model itself, enabling cross-task generalization.
- EFT as a 'practice phase' for general-purpose discovery agents that do not solve new problems from scratch.
Methodology
EFT fine-tunes open-source LLMs (2B to 9B parameters) on the Finch Collection dataset of evolutionary search trajectories. The models are evaluated on 22 held-out optimization tasks and further paired with test-time reinforcement learning.
Key Results
EFT models outperform base counterparts by 10.22% on average across 22 held-out tasks. With test-time RL, the model matches state-of-the-art on two circle-packing tasks and outperforms its base counterpart on the Erdős minimum-overlap problem.