GAE: Graph-Augmented Evolution for Scientific Discovery via Reinforcement Optimization
TLDR
GAE combines graph neural networks and reinforcement learning to guide LLM-based evolutionary search for automated scientific discovery, achieving state-of-the-art symbolic regression.
Reasoning
The paper presents a novel framework that addresses key limitations in LLM-driven evolutionary search through a three-pillar architecture, with strong empirical results on symbolic regression. However, the evaluation is limited to a single task, and the abstract does not demonstrate broader applicability or real-world experiments.
Read-first score
Read-first score 51.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 60.
Field roles
Rank sensitivity
Stability: volatile; rank range: 62.
Keyword Scores
Deep Analysis
Innovations
- Graph neural network that parses programs into typed computation graphs to produce structure-aware embeddings
- RL-optimized meta-controller that uses structure-aware embeddings for directed parent and mutation direction selection
- Online GRPO fine-tuning loop that adapts the LLM mutation operator at test-time using group-normalized evaluation rewards
Methodology
GAE integrates a relational GNN for program embeddings, an RL meta-controller that selects parents and mutation directions based on reward history, and an online GRPO loop that fine-tunes the LLM mutator. The framework is evaluated on symbolic regression for complex nonlinear oscillator systems, compared against static LLM-driven baselines.
Key Results
GAE consistently matches or outperforms static LLM-driven baselines and achieves state-of-the-art out-of-distribution performance on symbolic regression for nonlinear oscillators.