What it is
Mid-trains LLMs on evolutionary search traces so they learn reusable solution-improvement skills across optimization tasks instead of starting each task from scratch. Instead of leaving evolution skill in the search scaffold, EFT trains the model on search trajectories as next-mutation supervision, so smaller open LLMs internalize how to iteratively improve solutions across tasks.
Key results
- +10.22% avg over base models on 22 held-out tasks
- 355 training tasks vs 15 gives +14.1% avg held-out performance
- Matches SOTA on 2 circle-packing tasks with test-time RL
Discovery strategies learned from many optimization problems can transfer across tasks and make smaller open LLMs stronger mutation operators.