Hugging Face has introduced the Large Discovery Model (LDM), a recurrent architecture designed to tackle open-ended scientific discovery across diverse domains such as molecules, protein sequences, and computer programs. The work, detailed in a recent paper, addresses the challenge of optimizing expensive-to-evaluate objectives over vast, structured hypothesis spaces.
LDM couples a generative model—typically a large language model (LLM)—with a Bayesian non-parametric reward surrogate. The generative model proposes and refines candidate designs, while the surrogate predicts performance and quantifies uncertainty. This uncertainty-aware value guides candidate generation, refinement, and selection. Crucially, the discovery memory and surrogate are continually updated as new experimental observations arrive, enabling iterative improvement.
The researchers evaluated LDM on three scenarios: neural-network training, antibody design, and molecular optimization. Compared to LLM-only reflection and traditional statistical search, LDM achieved a 2.4× greater reduction in validation BPB, an 18.2% relative decrease in binding energy, and more than 60% relative gains in molecular multi-objective performance.
The paper positions LDM as a unified architecture that combines LLMs with Bayesian experimental design, with the potential for LLMs to learn experimental design through post-training. The initial release includes case studies as proof of concept, demonstrating effectiveness across diverse open-ended design problems.