Discovery Foundation ModelsFrom Solving Problems to Discovering Knowledge

What foundation models learn from

Foundation models have advanced in stages, each defined by what they learn from. They first learned from what humans wrote: large-scale pretraining turned them into broad stores of existing knowledge. They then began to learn from the outcomes of their own actions: reasoning, tool use, code execution, and agentic workflows driven by verifiable feedback.

In science, the same capabilities now support protein structure prediction, materials discovery, and weather forecasting, and AI co-scientists, AI-scientist systems, and self-driving laboratories increasingly connect model-generated hypotheses to experiments, mostly within goals and setups that people have chosen. This workshop asks what the next step looks like: foundation-model systems that take part in creating and validating new knowledge.

The ceiling: a research structure fixed by humans

Most current systems, however capable, begin from a research structure that people have already chosen, namely the problem, its representation, the objective, the tools, and the evaluator, and they search for a solution inside it. The limitation is not a lack of scientific knowledge or reasoning. Most training and evaluation pipelines reward competence only after this structure has been fixed, and more competence inside the structure does not by itself repair it.

A fixed representation cannot express a variable it omits, so a persistent failure may reflect a missing variable rather than a weak optimizer; a search objective cannot recover a property its evaluator ignores, so a benchmark can reward a proxy; and more samples cannot separate two mechanisms when the available observable does not distinguish them.

An automated workflow can therefore pursue a misframed question more efficiently without becoming any better at noticing the misframing. Scientific-agent evaluations make the gap visible: long research workflows run end to end while evidence integration and refutation-driven revision remain fragile. The stakes go beyond efficiency, because AI may already change which problems are pursued, not only how quickly they are solved.

From problem solving to discovery intelligence

We refer to the missing capability, changing the research structure itself rather than only working inside it, as discovery intelligence. In a discovery episode this structure is not fixed: the episode starts from a partially understood world and must produce both externally validated progress and a record of how the problem, representation, hypotheses, and evidence were revised.

Three transitions are still left to people:

• framing, turning observations and uncertainty into a problem worth pursuing;
• modeling, constructing the variables in which competing explanations become expressible and distinguishable;
• grounding and revision, choosing evidence that can change an explanation’s status and attributing failures to the right cause.

A DFM acts not on a fixed task but on an explicit, revisable research state: a linked record of the current problem, representation, hypotheses, evidence, and interventions, and of why each was revised. It is defined by what it can construct and revise, not by its architecture or degree of autonomy, and the unit that is evaluated and credited is the whole system, including its tools, validators, and human collaborators.

Why science

Science is where this capability can be learned and tested, because it is an environment that pushes back. Questions are underspecified, variables hidden, interventions costly, and outcomes delayed; above all, evidence arrives from experiments, simulations, and measurements that the model cannot rewrite after seeing the result.

These properties turn problem formulation, experiment design, failure attribution, and revision into observable decisions rather than qualities of a final answer, so they can be trained on directly and credited even when the outcome is negative.

Science also makes evaluator incompleteness visible. Benchmark leakage, simulator artifacts, and non-reproducible effects can all create apparent progress without stronger knowledge, so for a DFM the evaluator can itself become part of the research state when evidence suggests that it is misaligned. The same decisions arise in digital research, in simulators, and in physical laboratories; what changes across these settings is the cost and independence of the evidence.

Discovery is recursive

Validated findings update our understanding of the world and seed the next round of questions, and a discovery system can go further by revising its own skills, tools, simulators, memory, and evaluators. Self-improving agents that rewrite their own code and scaffolding, building on the theoretical Goedel machine, provide digital precedents.

Recursive scientific improvement carries a larger validation burden, because one update can change which problems are selected, which evidence is collected, and how success is judged: a benchmark generator that silently drops hard cases, or an evaluator that starts rewarding internal agreement, can make measured performance rise while real discovery ability narrows.

The find, test and improve agenda

The workshop organizes this question into a unified agenda for systems that find, test, and improve: find what is worth discovering (framing and modeling), test explanations against the world (grounding and revision), and improve the discovery process itself (learning and recursion). The challenges, topics, and program below follow this order.

Two dedicated tracks, Open Discovery Problems and Negative Results and Failure Analyses, bring problem owners from the sciences and well-attributed failures into the program, and a closing panel debates the questions this agenda leaves open. By bringing these communities together, we aim to help shape a roadmap toward AI systems that discover rather than only solve.

Tracks and page limits

Two further tracks reflect the theme of the workshop: Open Discovery Problems (up to 2 pages), in which scientists describe an unsolved problem together with the data and validation conditions needed to attack it, and Negative Results and Failure Analyses (up to 4 pages), which treats a well-attributed failure of an AI-driven discovery attempt as a contribution.

The tiny paper track welcomes less-than-full-conference work, such as an implementation and evaluation of an unpublished but simple idea, a modest but self-contained theoretical result, a follow-up experiment or re-analysis of a published paper, or a fresh perspective on an existing publication, and is meant to broaden participation by students, early-career researchers, and domain scientists; as ICLR requires, tiny papers must be primarily human-authored.

Non-archival

The workshop is non-archival, and papers already published at an archival venue, including the ICLR 2027 main conference, are not eligible.

Encouraged author disclosure

Authors are encouraged, but not required, to state briefly which parts of the research structure were given and which the system constructed or revised, which evidence supports each claim, and where humans intervened.

Review process

Reviewing is double-blind on OpenReview, with at least three reviews per full paper and two per shorter submission; we expect 80–150 submissions and will recruit about 160 reviewers, so that no reviewer handles more than three papers even if every submission is a full paper, and each paper is assigned both ML and domain-science reviewers.

• Double-blind, on OpenReview.
• At least three reviews per full paper; two per shorter submission.
• Expected submissions: 80 to 150. Reviewers to recruit: about 160.
• No reviewer handles more than three papers.
• Every paper is assigned both an ML reviewer and a domain-science reviewer.

Conflicts of interest

No organizer handles a submission from their own organization; submissions from the organizers’ groups or PhAI Labs are handled by a senior program-committee member outside these groups and are not eligible for awards.

Use of AI systems

AI systems will not serve as authors or reviewers: AI-generated submissions will not be accepted, AI-assisted writing is permitted if clearly disclosed in accordance with the ICLR 2027 policy, and reviewers may not use AI tools to write their reviews or upload submissions to external AI services.

Acceptance volume and presentation

We expect to accept 40–60 papers. All accepted papers will be presented as posters, split by theme across the two poster sessions, with the strongest full and tiny papers selected for contributed talks, the strongest negative results for spotlights, and the strongest open problems for pitches; outstanding submissions are recognized in the closing awards session.

Format

We propose a full-day, in-person workshop combining invited talks, contributed presentations, posters, and a panel debate to balance intellectual depth with structured interaction. The program will feature eight invited talks, four contributed talks, three spotlights from the Negative Results and Failure Analyses track, a pitch session for the Open Discovery Problems track, two 1-hour poster sessions integrated with coffee breaks, and a closing panel debate with four of the invited speakers. The invited talks follow the agenda: find in the morning, test in the early afternoon, and improve in the late afternoon. In total, the poster sessions, problem pitches, and panel debate give about three hours of structured interaction. An awards session will conclude the workshop.

• 8 invited talks.
• 4 contributed talks, 10 minutes each.
• 3 spotlights from the Negative Results and Failure Analyses track.
• 1 pitch session for the Open Discovery Problems track.
• 2 one-hour poster sessions, integrated with coffee breaks.
• 1 closing panel debate with four of the invited speakers.
• About 3 hours of structured interaction in total, and an awards session to conclude.

The closing panel debate

The workshop closes with a panel of four invited speakers, van der Schaar, Stanley, Dieng, and Liu, organized around three questions on which we expect them to disagree: how far a discovery system should be allowed to revise its own evaluator, and what must stay outside its self-modification; whether many automated runs are stronger evidence of discovery ability than a few well-audited investigations; and how credit for a discovery should be divided between the model and the humans, tools, and laboratories around it. The panel is moderated by one of the organizers, who also take audience questions; organizers give no talks at the workshop.

The three questions, in full:

1. How far a discovery system should be allowed to revise its own evaluator, and what must stay outside its self-modification.
2. Whether many automated runs are stronger evidence of discovery ability than a few well-audited investigations.
3. How credit for a discovery should be divided between the model and the humans, tools, and laboratories around it.

Awards

We will present up to three awards: Best Paper, Best Open Discovery Problem, and Best Grounded Discovery, the last for the submission whose claims rest most convincingly on evidence the system cannot rewrite, such as wet-lab experiments, independent replication, or hidden-test and held-out validation in digital settings, whether the outcome confirms or refutes the original hypothesis. Winners are presented in the contributed talks, spotlights, or pitches.

Up to three is a ceiling, not a promise. The proposal names no prize amounts.

Why these speakers

Each speaker’s work is a building block of the agenda, and the invited talks follow the find-test-improve order of the day. Stanley opens find with open-ended search for what is worth discovering, now tested with Lila Sciences’ physical laboratories. Dieng follows with Vendi scores and diversity-driven experimental design, which steer molecular and materials discovery toward unexplored regions. Ji brings SciMON, which generates scientific ideas and optimizes them for novelty against the literature. Liu closes find with multimodal generative models that connect language with molecules and proteins, and with clawRxiv, which raises the question of how research written by AI agents should be judged.

In test, Ma builds computer-using agents that run laboratory workflows (Orion) and benchmarks that check how well agents carry out real scientific work (ScienceBoard). Bian builds reasoning-capable foundation models for atomic systems and LLM agents for science, which tie a model’s reasoning to physical structure. Van der Schaar, whose lab studies networks of research agents and human experts that jointly generate, critique, and refine hypotheses, speaks to how explanations are revised against evidence and how credit is divided between the model and the people around it. Clune closes the invited talks and anchors improve with The AI Scientist and the Darwin Goedel Machine.

Track distribution of the eight proposed talks

Interaction and access

ICLR 2027 workshops are held in person. To engage those who cannot attend, invited talks and the panel will be recorded where ICLR provides recording and the speakers agree; otherwise we will post speaker-provided videos or slides on the workshop website. Accepted papers, posters, and slides will be posted on OpenReview and the workshop website, and talk titles will be public before the conference. Within four weeks of the workshop we will publish on the website the curated list of open discovery problems with their validation conditions, together with a short report on the panel debate. Authors or speakers who cannot travel, for example because of visa delays, can present through a pre-recorded video, with their poster displayed by an organizer; if an invited speaker withdraws, the slot goes to an additional contributed talk. Questions for the panel will also be collected through an online form during the day.

Expected audience and outreach

We anticipate 150–200 sustained in-person participants, with peak attendance of 250–350. This estimate is based on comparable ICLR workshops and on the more than 250 submissions received by the ICLR 2026 Lifelong Agents workshop, co-organized by members of our team. We will promote the workshop through a dedicated website, social media, OpenReview, mailing lists in machine learning and in computational biology, chemistry, and materials science, and the organizers’ and speakers’ networks. For the Open Discovery Problems track, we will reach out directly to self-driving-laboratory consortia and domain-science communities, since real problem owners are what make the track useful.

Inclusion and broader engagement

A central part of our outreach is active engagement with affinity groups, including WiML, LatinX in AI, Black in AI, Queer in AI, and {Dis}Ability in AI, as well as universities in the San Francisco Bay Area. We will host mentoring tables at lunch, open at least one contributed talk to the tiny paper track, and invite students to join the program committee as reviewers paired with senior reviewers.

In inviting speakers the organizers drew on a candidate list that includes researchers from groups underrepresented in machine learning, and they will give priority to gender balance and to these groups if the speaker list changes after acceptance and when recruiting the program committee. Access measures named in the proposal are online materials, pre-recorded presentations for those who cannot travel, the tiny paper track, and the allocation of complimentary registrations to students and early-career authors.

Funding support