SMCEvolve: Principled Scientific Discovery via Sequential Monte Carlo Evolution
TLDR
SMCEvolve uses Sequential Monte Carlo to guide LLM-driven program evolution for principled scientific discovery with convergence guarantees.
Reasoning
The paper introduces a novel SMC-based framework that provides theoretical guarantees and outperforms existing systems on diverse benchmarks. Its strength lies in principled design and finite-sample analysis, but the abstract lacks details on limitations or comparisons to non-LLM methods.
Read-first score
Read-first score 68.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 71.
Field roles
Rank sensitivity
Stability: volatile; rank range: 38.
Keyword Scores
Deep Analysis
Innovations
- Recasting program search as sampling from a reward-tilted target distribution using Sequential Monte Carlo
- Adaptive parent resampling
- Mixture of mutation with acceptance
- Automatic convergence control
- Finite-sample complexity analysis bounding LLM-call budget for target approximation error
Methodology
SMCEvolve frames program evolution as sampling from a reward-tilted distribution using a Sequential Monte Carlo sampler, incorporating adaptive parent resampling, a mixture of mutation with acceptance, and automatic convergence control. It includes a finite-sample complexity analysis bounding LLM-call budget for target approximation error, and is evaluated on math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks.
Key Results
SMCEvolve surpasses state-of-the-art evolving systems across math, algorithm efficiency, symbolic regression, and end-to-end ML research benchmarks while using fewer LLM calls under self-determined termination.