Evaluation-driven Scaling for Scientific Discovery
Abstract: Language models are increasingly used in scientific discovery to generate hypotheses, propose candidate solutions, implement systems, and iteratively refine them. At the core of these trial-and-error loops lies evaluation: the process of obtaining feedback on candidate solutions via verifiers, simulators, or task-specific scoring functions. While prior work has highlighted the importance of evaluation, it has not explicitly formulated the problem of how evaluation-driven discovery loops can be scaled up in a principled and effective manner to push the boundaries of scientific discovery, a problem this paper seeks to address.
We introduce Simple Test-time Evaluation-driven Scaling (SimpleTES), a general framework that strategically combines parallel exploration, feedback-driven refinement, and local selection, revealing substantial gains unlocked by scaling evaluation-driven discovery loops along the right dimensions. Across 21 scientific problems spanning six domains, SimpleTES discovers state-of-the-art solutions using gpt-oss models, consistently outperforming both frontier-model baselines and sophisticated optimization pipelines.
Key Achievements of SimpleTES
SimpleTES has demonstrated remarkable advancements in several scientific applications. Here are some notable achievements:
- Sped up the widely used LASSO algorithm by over 2x.
- Designed quantum circuit routing policies that reduce gate overhead by 24.5%.
- Discovered new Erdos minimum overlap constructions that surpass the best-known results.
Feedback-driven Learning and Generalization
Beyond novel discoveries, SimpleTES produces trajectory-level histories that naturally supervise feedback-driven learning. When models are post-trained on successful trajectories, they not only improve efficiency on seen problems but also generalize to unseen problems, discovering solutions that base models fail to uncover. This dual capability enhances the models’ adaptability and effectiveness in tackling diverse scientific challenges.
Conclusion
Together, our results establish effective evaluation-driven loop scaling as a central axis for advancing LLM-driven scientific discovery. The SimpleTES framework provides a simple yet practical approach for realizing these gains, paving the way for further innovations in scientific research. As the integration of AI in scientific discovery continues to evolve, frameworks like SimpleTES will play a crucial role in harnessing the full potential of language models, ultimately leading to groundbreaking discoveries and advancements across various scientific domains.
The implications of this research extend beyond individual projects; they represent a shift in how scientific problems can be approached and solved using AI. By emphasizing evaluation-driven strategies, researchers can push the envelope of what is achievable, fostering an environment where scientific discovery is accelerated and enhanced through intelligent feedback mechanisms.