Evaluation-driven Scaling for Scientific Discovery
Summary
SimpleTES framework scales evaluation-driven discovery loops across 21 scientific problems, yielding 2× speedups on LASSO, 24.5% quantum gate reductions, and new Erdos constructions while enabling trajectory-level model post-training.
View Cached Full Text
Cached at: 04/22/26, 06:17 AM
Paper page - Evaluation-driven Scaling for Scientific Discovery
Source: https://huggingface.co/papers/2604.19341 Authors:
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
,
Abstract
SimpleTES framework scales evaluation-driven discovery loops for scientific problems, achieving state-of-the-art results across multiple domains through parallel exploration and feedback-driven refinement.
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 howevaluation-driven discovery loopscan 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 combinesparallel exploration,feedback-driven refinement, and local selection, revealing substantial gains unlocked by scalingevaluation-driven discovery loopsalong 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. Particularly, we sped up the widely usedLASSO algorithmby over 2x, designedquantum circuit routingpolicies that reduce gate overhead by 24.5%, and discovered newErdos minimum overlap constructionsthat surpass the best-known results. Beyond novel discoveries, SimpleTES produces trajectory-level histories that naturally supervise feedback-driven learning. When post-trained on successful trajectories, models not only improve efficiency on seen problems but also generalize to unseen problems, discovering solutions that base models fail to uncover. Together, our results establish effective evaluation-driven loop scaling as a central axis for advancingLLM-driven scientific discovery, and provide a simple yet practical framework for realizing these gains.
View arXiv pageView PDFProject pageAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2604.19341 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2604.19341 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2604.19341 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Scientific discovery as meta-optimization: a combinatorial optimization case study
This paper proposes formalizing scientific discovery as a meta-optimization problem where LLMs generate and aggregate objective functions via correlation-weighted voting, applied to 3-SAT algorithm discovery using digital MemComputing, achieving a 67x speedup on large instances.
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
This paper introduces AutoTTS, an environment-driven framework that automates the discovery of test-time scaling strategies for LLMs by formulating it as controller synthesis. It demonstrates improved accuracy-cost tradeoffs on mathematical reasoning benchmarks with minimal computational overhead.
TEMPO: Scaling Test-time Training for Large Reasoning Models
TEMPO introduces a test-time training framework that alternates policy refinement with critic recalibration to prevent diversity collapse and sustain performance gains in large reasoning models, boosting AIME 2024 scores for Qwen3-14B from 42.3% to 65.8%.
Saturating Scaling Laws for Equational Discovery: A Phenomenology of Growth Dynamics in Three Toy Substrates with Two Real-World Replications
This paper investigates growth dynamics in deterministic equational discovery across three toy substrates and two real-world replications, finding substrate-conditional saturating power-law scaling.
LEAP: Trajectory-Level Evaluation of LLMs in Iterative Scientific Design
The paper introduces LEAPBench, a 55-task framework for trajectory-level evaluation of LLMs in iterative scientific design, revealing that outcome-based scoring misses efficiency gains and that domain-agnostic prompting can outperform domain-aware prompting in matching published best designs.