Cost-Efficient LLM-Driven Scientific Discovery

Top: performance versus API cost on mathematical discovery. Bottom: equation discovery error over 500 evaluations.

LLM-driven evolutionary search has become a powerful tool for scientific discovery: models repeatedly propose candidate solutions, which are evaluated and refined over many iterations. Relying on frontier models for every step is expensive, while lightweight models are cheaper but less capable on their own. This project studies how to coordinate models of different strengths so that discovery stays both effective and affordable.

In SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery (arXiv preprint), we explore with lightweight models and consolidate with frontier reasoning. SpecEvo lets parallel lightweight Speculators propose candidates, an Advisor distill the search trajectory into reusable guidance, and a frontier Navigator intervene only at sparse checkpoints, so that model coordination adapts as the search progresses. Across 176 tasks in four domains, it is competitive with frontier-model baselines on mathematical discovery and systems optimization at up to 80% lower API cost, achieves a 3.8% higher mean normalized score than the strongest baseline on CO-Bench combinatorial optimization, and attains the lowest NMSE in all eight LSR-Synth equation discovery settings among the main evolutionary baselines. The SpecEvo repository provides the implementation.

Ha Minh Hieu
Ha Minh Hieu
SRI Research Intern (May ‘26)

Research Intern

Tongyao Zhu
Tongyao Zhu
IPP Doctoral Student (Jan ‘23; SEA)
Co-Supervised by Qian Liu.
Research Achievement Award (RAA, ‘25)

PhD Candidate January 2023 Intake

Xuan Long Do
Xuan Long Do
A*STAR Doctoral Student (Aug ‘23)
Co-Supervised by Kenji Kawaguchi and Nancy F. Chen
Research Achievement Award (RAA, ‘24); Dean’s Graduate Research Excellence Award (DGA, ‘26)

PhD Candidate August 2023 Intake

Min-Yen Kan
Min-Yen Kan
Associate Professor

WING lead; interests include Digital Libraries, Information Retrieval and Natural Language Processing.