Large Discovery Model Turns Scientific Search into a Feedback Loop

Scientific discovery is rarely a one-shot generation problem. Researchers propose a hypothesis, test it, learn from the result, and then decide what to try next. This loop becomes especially difficult when the search space is enormous and structured, while each evaluation is expensive. Examples include finding better neural-network training programs, designing antibodies with desired properties, and optimizing molecules against several objectives at once.

Large Discovery Model, or LDM, is presented as a unified architecture for this type of open-ended search. The paper starts from a practical limitation of generative models, including large language models: their likelihood scores are not reliable substitutes for the true objective. A model may assign high probability to a candidate because it looks familiar or syntactically plausible, not because it performs well in an experiment. Its self-assessment is also not necessarily a calibrated measure of epistemic uncertainty, particularly for candidates outside the observed data distribution.

LDM connects two complementary components:

• A generative model proposes and refines candidates. It can create new designs or iteratively improve existing ones.
• A reward surrogate estimates performance. Built with a Bayesian non-parametric approach, the surrogate predicts the likely outcome of a candidate while also representing uncertainty.
• An uncertainty-aware value guides decisions. Candidate generation, refinement, and selection are informed by both predicted reward and uncertainty, balancing exploitation of promising regions with exploration of less familiar ones.
• Memory is updated after every observation. New experimental results are added to the discovery memory and used to update the surrogate, shaping subsequent rounds of search.

The important conceptual shift is that the generative model is no longer treated as the final judge. It serves as a proposal engine inside a larger empirical loop. The system does not need to know the complete objective function in advance. Instead, it gradually focuses the search by comparing predictions with external observations. The surrogate acts as the bridge between an abstract candidate and its measured scientific value, helping allocate a limited experimental budget more efficiently.

The authors evaluate LDM in three settings: neural-network training, antibody design, and molecular optimization. According to the supplied abstract, LDM outperforms LLM-only reflection and traditional statistical search across these domains. It achieves a 2.4-times greater reduction in validation BPB, an 18.2% relative decrease in binding energy, and more than 60% relative gains in molecular multi-objective performance. Because the tasks span programs, proteins, and molecules, the results are intended to support the generality of the generation–evaluation–learning loop rather than a single domain-specific trick.

The work should nevertheless be read as an architecture and proof-of-concept evaluation, not as evidence that experiments can already be removed from the process. Its performance depends on the quality of experimental observations, the calibration of surrogate uncertainty, and the suitability of candidate representations and objectives. A promising direction is to train discovery systems over many rounds of real feedback, allowing them to learn not only how to generate candidates, but also how to choose the most informative next experiment.

For AI for Science, LDM highlights a broader lesson: expressive generation is only the beginning. In open-ended discovery, the more consequential capability is deciding what to test next when the model is uncertain and the cost of being wrong is high.

Source: Hugging Face Daily Papers