Large Discovery Model Adds Bayesian Surrogate to Guide LLM Search

Scientific discovery has become one of the AI industry's favourite generative-model pitches, but the team behind the Large Discovery Model argues the pitch is only half the machine. In a preprint posted to arXiv, Zhongwei Yu and collaborators propose coupling a generative model with a Bayesian non-parametric reward surrogate that scores candidates and, crucially, quantifies its own uncertainty about them.

The setup runs as a loop. The generator proposes and refines designs, the surrogate predicts performance, and an uncertainty-aware value decides what to try next. "The discovery memory and the surrogate model are continually updated as each new experimental observation arrives," the abstract says, framing the architecture as an online learner rather than a one-shot ranker.

Across three test domains — neural-network training, antibody design, and molecular optimisation — the authors report "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" versus LLM-only reflection or traditional statistical search.

The broader claim is more tentative. The abstract offers that LDM "could serve as a general-purpose discovery engine for effective search over open-ended hypothesis spaces" — a hedge, not a proof. The abstract publishes no per-baseline breakdowns, no compute budgets, no wet-lab validation for the antibody result, and does not name the specific LLMs paired with the surrogate.