Discovery Foundation Models: Toward Open-Ended Discovery Intelligence
Ling Yang, Zhenfei Yin, Yingcheng Wu
cs.CL
2026-09-15
A framework paper defines discovery as seven coupled operations, instantiates Zetema, and grounds the loop in GALILEO wet-lab work, without standard benchmark scores.
Most scientific AI still solves a structure that people already chose. The question, variables, objective, tools, and evaluator arrive as given. Models can search, reason, and run long workflows inside that box. They are much weaker when progress requires changing the box: reformulating the unknown, swapping a representation, or designing an intervention that makes two explanations disagree.
Long research traces can succeed as workflows while evidence integration, refutation-driven revision, and transfer stay fragile. The paper names the missing target Discovery Intelligence: a system that takes part in how new problems, representations, explanations, and knowledge get made, rather than only answering tasks that already have a score.
A Discovery Foundation Model (DFM) is a model system, not a single network. It acts on a revisable research state through seven coupled capabilities:
The evaluated unit is D = (policy, research state, memory, tools, environment, validators, humans). A human’s decisive reformulation has to be logged, not credited to the model.
Zetema is the instantiation: explicit state, branching and rollback, a Research World Model that gates costly actions, Dry-Lab reasoning tied to external evidence, and promotion of validated experience into Discovery Skills. Training aims at intermediate research decisions. Evaluation scores both externally validated progress in the current episode and later discovery behavior under matched resources.
The empirical anchor is GALILEO, a therapeutic-peptide loop. Dry-Lab target nomination and sequence design meet robotic synthesis and phenotyping; physical outcomes re-enter the decision loop. On LRRC8C-high organoids the measured IC50 is 18.922 μM and the working story moves from generic membrane toxicity toward osmolyte/redox and innate signaling. On SLC25A1 the IC50 is 0.736 μM and the story tightens toward citrate-export-dependent metabolic collapse. Across five optimization rounds, feedback is distilled into a transferable Amphiphilic Balance Grammar.
This is a framework paper. There is no standard leaderboard and no head-to-head of Zetema against existing scientific agents under one protocol. The numbers that can be quoted come from GALILEO: two target branches with organoid IC50 values, five design rounds, and one portable design rule. The authors keep the claim narrow. The case shows that physical evidence can change later scientific actions, including reopening or killing branches and rewriting design policy. It does not show that a general-purpose DFM already exists.
A comparison table is the other concrete artifact. Foundation models for science, reasoning models, scientific agents, search systems, and autonomous labs usually take the problem as given and do not treat skill transfer as defining. A DFM requires constructed problems, revisable representations, required intervention and external grounding, and required transfer.
The useful product is a checklist. A system that “does discovery” should be able to point to which object it changed, where the evidence came from, and whether a skill transferred under controls. For people building research agents, scoring only the final answer hides the decisions that actually move a program. The public repo DFM-Plans is a collaboration and planning entry point, not a trained general discoverer.
Read it as an agenda, not as a model card.
GALILEO’s domain, molecular modality, and lab interfaces remain highly structured; cross-domain transfer is still open. The training and evaluation protocol is specified at length, but the paper does not report a large run under that protocol. Electrophysiology, metabolomics, and in-vivo work can stay human-executed, so the capability claim belongs to the coupled system. Governance discussion of provenance, scoped authority, and recursive update risk is qualitative. Using one peptide loop as empirical grounding for general discovery intelligence is a long extrapolation.