Paper

Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search

Zhongwei Yu, Yan Song, Xue Yan, Anjie Liu, Xingyu Lu, Yihang Chen, Huichi Zhou, Siyuan Guo

An LLM-plus-surrogate search loop that uses uncertainty-aware acquisition to steer generation, refinement, and expensive evaluations in open-ended design spaces.

arXiv ↗Aug 16, 2026Code ↗

multimodal
bayesian optimization
scientific discovery
inverse design

Pipeline diagram showing slow/fast learning loop with LLM reservoir and GP model

What it is

Couples an LLM proposal model with a Bayesian reward surrogate to run uncertainty-aware search over open-ended design spaces for scientific discovery. Instead of trusting LLM likelihood or using BO on a fixed space, it combines LLM candidate generation with a Bayesian non-parametric surrogate whose uncertainty guides proposal, refinement, and evaluation as observations arrive.

Key results

2.4× larger val-BPB reduction than LLM-only reflection on AutoResearch
18.2% lower mean binding energy than LLM-only after 200 steps
62.4% higher molecular Pareto hypervolume than LLM-only; 63.1% over BO

Shows that open-ended generative search becomes much more effective when empirical reward modeling and epistemic uncertainty drive which candidates to pursue.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

What it is

Key results

One useful ML email a week.

More from Aug 17 – Aug 23

Worth your time, not all of your time.