Scientific Agents
Scientific foundation models capture patterns in scientific data, but a useful prediction is only one step in a scientific investigation. We focus on scientific agents: systems that use foundation models to formulate and revise hypotheses, select and execute tool-mediated actions, interpret evidence, and decide what to investigate next.
From scientific foundation models to agents that formulate hypotheses, choose experiments, and learn from evidence.
The third Workshop on Foundation Models for Science brings together machine learning researchers, domain scientists, and scientific infrastructure developers across scientific computing, physical and life sciences, and computational mathematics. We ask:
- How should scientific agents be designed and trained?
- How should they use scientific knowledge and feedback?
- What evidence establishes that they make reliable scientific progress?
Scientific agents connect general reasoning models with specialized models, simulators, numerical solvers, and analysis software. We welcome single-agent systems, multi-agent teams, search procedures, and human-guided workflows. Full autonomy is not required, and rigorous studies of failure are central to the workshop. We seek reproducible gains in validated results per unit of compute, data, and human effort against explicit baselines.
Building on our NeurIPS 2024 and ICLR 2026 editions, this workshop shifts the focus to adaptive scientific investigations: how models are composed, actions are selected, and new evidence changes subsequent decisions. Each prior edition received 100 submissions; our 2027 planning estimates are 150-250 attendees and 100-150 submissions.
Scope and Open Problems
- Scientific tool use and composition of foundation models. Connect language and multimodal models with domain-specific predictors, neural operators, numerical solvers, symbolic tools, and scientific databases. Study action representations, tool selection, interface validation, and uncertainty propagation.
- Experimental design and learning from feedback. Choose measurements, simulations, and interventions under limited budgets through active learning, Bayesian optimization, causal experimental design, reinforcement learning, multi-fidelity simulation, and adaptive stopping.
- Multi-agent collaboration and human scientific judgment. Understand when specialist roles, debate, and independent criticism improve scientific reasoning. Address correlated errors, evidence-grounded disagreement, human oversight, and comparisons with equally resourced single-agent systems.
- Long-running investigations, memory, and research workflows. Maintain coherent records of hypotheses, datasets, code, results, and rejected explanations. Develop persistent scientific memory, provenance, experiment search, reproducible workflows, and recovery from failed tools.
- Discovery through executable evaluation. Discover algorithms, numerical methods, mathematical constructions, and scientific software through program search and executable feedback. Study evaluator design, inference-time compute allocation, hidden-test generalization, and specification errors.
- Reliable evaluation and responsible scientific autonomy. Distinguish workflow completion, replication, hypothesis quality, and independently validated discovery. Explore contamination-aware benchmarks, uncertainty calibration, negative results, robustness, auditable tool execution, and human oversight.
We welcome methodological, theoretical, systems, and application contributions. Our discussions will inform a reporting checklist, a shared evaluation agenda, and a directory of contributed scientific tools, datasets, and reproducible workflows.
Confirmed Speakers / Panelists (A-Z by Last Name)
All nine invitees have confirmed participation, agreed to be listed, and verified that they appear as invited speakers or panelists in no more than two ICLR 2027 workshop proposals.
Ian T. Foster
Professor, University of Chicago
Distinguished Fellow, Senior Scientist, MCS Division, Argonne
Shirley Ho
Group Leader, Cosmology X Data Science, CCA, Simons Foundation
Professor, Department of Physics & Center for Data Science, NYU
Bo Li
Abbasi Associate Professor, University of Illinois Urbana-Champaign
Ge Liu
Assistant Professor, University of Illinois Urbana-Champaign
Michael Mahoney
Adjunct Professor, University of California, Berkeley
Vice President, International Computer Science Institute (ICSI)
AI Initiative Research Lead, Scientific Data Division, Lawrence Berkeley National Laboratory
Yizhou Sun
Professor, UCLA
Call for Papers
We welcome new and ongoing work on scientific agents, including preliminary ideas, focused experiments, replications, negative results, tools, and demonstrations.
Regular papers: up to 8 pages. Short papers: 2-4 pages. References and appendices are excluded from these limits. Both tracks share deadlines, receive three human reviews per paper, and are eligible for contributed talks and posters. Authors select the track appropriate to their contribution's scope. The short track welcomes late-breaking and early-stage work and lowers the submission barrier for budding researchers and under-represented communities.
The workshop is non-archival. Accepted papers will be publicly available on OpenReview.
Read the submission guidelines. The submission portal and template link will be announced here.
Important Dates
All deadlines are at 23:59 Anywhere on Earth (AoE) and apply to both tracks.
- Abstract registration: January 25, 2027
- Final paper submission: February 1, 2027
- Internal decisions: February 23, 2027 (not author notification)
- Author notification: February 26, 2027
- Accepted papers public on OpenReview and imported to iclr.cc: February 26, 2027
- Workshop: April 29 or 30, 2027, as assigned by ICLR
Workshop Schedule
The one-day, in-person program starts at 09:00 (9:00 a.m.) and ends at 17:00 (5:00 p.m.), with seven invited talks and a panel drawn from the nine confirmed invitees. Times are local to the venue; ICLR assigns the workshop day and room.
| Local time | Session |
|---|---|
| 09:00 - 09:05 | Opening remarks |
| 09:05 - 09:35 | Invited talk 1 |
| 09:35 - 10:05 | Invited talk 2 |
| 10:05 - 10:50 | Poster session I |
| 10:50 - 11:20 | Invited talk 3 |
| 11:20 - 11:50 | Invited talk 4 |
| 11:50 - 12:20 | Contributed talks 1-3 |
| 12:20 - 13:05 | Lunch |
| 13:05 - 13:35 | Invited talk 5 |
| 13:35 - 14:05 | Invited talk 6 |
| 14:05 - 14:50 | Poster session II |
| 14:50 - 15:20 | Invited talk 7 |
| 15:20 - 16:00 | Debate panel |
| 16:00 - 16:30 | Contributed talks 4-6 |
| 16:30 - 16:50 | Challenge breakout tables |
| 16:50 - 17:00 | Synthesis and closing |
Each invited talk lasts 30 minutes: 25 minutes of presentation and 5 minutes of questions. Contributed talks include 8 minutes of presentation and 2 minutes of questions. Two 45-minute poster sessions provide 90 minutes for contributed work and breaks between talks. All accepted papers have a dedicated poster slot, with authors presenting in two balanced groups.
Seven of the nine invitees will give talks; the other two will participate only in the 40-minute panel, moderated by Wuyang Chen. One panelist will present the case for multi-agent collaboration; the other will examine whether an equally resourced single agent can achieve the same gains. These are assigned debate roles. Comparisons will match compute, tools, data, and human assistance.
One week before the workshop, we will circulate a draft checklist item and comparison template for each of the six challenges, with fields for task, baseline, budget, metric, and failure criterion. The 20-minute breakout session is for refining these drafts: 3 minutes to review the starting material, 12 minutes to discuss changes, and 5 minutes to record one priority, disagreements, and open fields in a shared document. A fully specified experiment is not required during the session.
Table facilitators: Wuyang Chen leads challenge 1 (model composition) and challenge 5 (executable evaluation); N. Benjamin Erichson leads challenge 2 (experimental design); Yuyang (Bernie) Wang leads challenge 3 (collaboration); Mayank Jauhari leads challenge 4 (memory and workflows); Mansi Sakarvadia leads challenge 6 (reliable evaluation). Chen's two tables will be adjacent, and he will spend ten minutes with each. Each table selects a scribe during the opening three minutes to maintain the shared notes throughout the session and afterward.
The 10-minute synthesis and closing session allows one minute per table to report its priority and four minutes for shared priorities and next steps. Table notes remain editable for one week after the workshop. Organizers will consolidate the reporting checklist and evaluation agenda for publication within four weeks.
Organizers
Wuyang Chen
Assistant Professor, Simon Fraser University
Ben Erichson
Staff Scientist, LBNL; Leader, Deep Learning Group, ICSI
Mayank Jauhari
Applied Scientist, Amazon
Mansi Sakarvadia
Ph.D. Student, University of Chicago
Yuyang (Bernie) Wang
Member of technical staff, NTT DATA AIVista
Workshop Proposal Review
The required external proposal-review committee nomination has been submitted through OpenReview.
Participation and Access
We welcome researchers across disciplines, career stages, and institutional backgrounds. The short-paper track, poster sessions, written questions, and facilitated discussions create opportunities for newcomers and early-career researchers to participate.
The event is planned in person. Subject to permissions and recording arrangements, we will share slides, recordings or speaker-provided videos, posters, and a discussion synthesis online. Authors facing exceptional travel barriers may provide a prerecorded presentation and nominate an in-person poster presenter.
Sponsorship and Participant Support
We are seeking support for students and under-resourced participants. Sponsorship and travel support are not yet confirmed. Please contact foundationmodelscience@gmail.com to discuss supporting the workshop.