EFT: Internalizing Evolutionary Search Ability into Small Language Models
> Paper: *Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks* > arXiv: 2606.29082 | Young-Jun Lee et al. (University of Minnesota) > Dataset: Finch Collection (156K trajectories, 10 domains, 371 tasks)
The Problem with Existing Methods: Capability Lives in the Scaffolding, Not the Model
LLM + evolutionary search has achieved notable breakthroughs in optimization tasks: GPU kernel design, mathematical conjecture proofs, scientific law discovery... But existing approaches share a fundamental flaw:
The evolutionary capability lives in the "scaffolding," not in the "model."
Every time the system encounters a new problem, the model starts from zero—search experience is discarded after use. The model itself cannot iteratively optimize, cannot decide which part to mutate, cannot determine when to roll back. All of these capabilities are provided by the external search framework.
EFT asks a deeper question: what if the model learned to evolve on its own?
The Core Idea of EFT: Evolutionary Trajectories as Supervision Signal
Rather than having the model search from scratch at test time, EFT lets it "practice" evolution during training.
Training Data: The Finch Collection
- 156K evolutionary search trajectories
- 10 domains: mathematics, competitive programming, GPU kernels, scientific law discovery, combinatorial puzzles...
- 371 optimization tasks
- Generated using Qwen3.5-397B running the OpenEvolve framework, with erroneous trajectories filtered out to retain high-quality samples
- Identify problems in the current solution
- Decide where and how to mutate
- Evaluate directions for improvement
- Know when to roll back
- Circle-packing task: matches SOTA-level performance
- Erdős minimum overlap problem: surpasses the base-model counterpart
- Paper: https://arxiv.org/abs/2606.29082
- Authors: Young-Jun Lee, Seungone Kim, et al. (University of Minnesota)
Training Paradigm
Evolutionary search trajectories are converted into supervision signals for training open-source models of 2B–9B parameters to:
This doesn't teach the model "the answer to this problem"—it teaches the model "how to find the answer."
Experimental Results: Small Models Can Also Make Discoveries
Cross-Task Generalization (22 Held-Out Tasks)
EFT fine-tuned models improve over base models by an average of 10.22%.
The key point is not absolute scores but cross-domain transfer—the model transfers optimization strategies from recommender systems to competitive programming, and GPU kernel design experience to mathematical conjectures. This emergent transfer suggests the model has genuinely internalized some general "discovery intuition."
Combining with Test-Time RL
EFT model + test-time reinforcement learning:
This shows EFT is not a replacement for search frameworks, but rather makes the model a better search participant—it can be embedded directly into existing frameworks as a mutation operator (with frozen weights), or further adapted with test-time learning.
Why This Matters
1. Lower barrier to entry — no need for closed-source large models (GPT-4, Claude); 2B–9B open-source models can perform high-quality evolutionary search
2. Internalized capability — evolutionary search shifts from an "external tool" to a "model instinct"; new tasks no longer start from zero
3. Reproducible dataset — Finch Collection is public: 156K trajectories covering 371 tasks, ready for follow-up research
4. Orthogonal to existing methods — EFT models can plug into any search framework and can be stacked with test-time RL without conflict
An Analogy
Traditional methods are like handing a student a brand-new dictionary (the search framework) for every exam—the student (the model) doesn't know how to use a dictionary and just reads along.
EFT is like letting the student first spend extensive time practicing dictionary lookup, note-taking, and pattern summarization. On exam day the dictionary is still available, but the student already knows what to look up, how to look it up, and when to try a different approach.