About

AI for Science is moving beyond individual models toward systems that participate in iterative scientific discovery. As AI systems generate hypotheses, use scientific tools, design experiments, interpret observations, and adapt their next actions based on feedback, a central challenge is how to engineer environments that compose these capabilities into reliable scientific workflows.

Our ICLR 2027 workshop, Scientific Workbenches for Discovery, brings together AI researchers, tool builders, and experimental scientists to develop a shared research agenda for reliable, interactive, and iterative scientific discovery.

We use the term scientific workbench for an integrated system of models, tools, data, experimental interfaces, orchestration mechanisms, and evaluators that supports an iterative loop:

Hypothesis → Action / Experiment → Observation → Feedback → Revision

Main themes

Loop Engineering: Design, optimize, and strengthen scientific interaction cycles across repeated rounds of reasoning, experimentation, and feedback.
Harnesses & Orchestration: Coordinate heterogeneous scientific tools, data sources, and computational resources into executable workflows.
Interactive Benchmarks & Evaluation: Assess scientific systems over trajectories of decisions, analyses, and experiments, rather than only static answers.
Scientific Environments: Build the computational and physical environments in which AI systems act, experiment, and receive feedback.

Many components of scientific workbenches are emerging in isolation, with different interfaces, assumptions, and objectives. This workshop aims to bring these directions together. Attendees will gain a clearer framework for scientific workbenches, research principles and open problems for loop engineering and scientific harnesses, new directions for interactive evaluation, and connections across machine learning and the sciences.

Important Dates (Anywhere on Earth)

Submission deadline (papers and competition proposals): February 1, 2027 AoE
Acceptance notifications: February 26, 2027 AoE
Workshop date: To be announced

Submissions

We invite submissions in two research tracks:

Original Research Track: Original studies using AI to tackle problems across scientific disciplines.
Position Track: Perspectives on current progress, open questions, and concerns in AI scientists research.

Research papers should contain 4-8 pages of main text, use the ICLR style file, and may include unlimited appendices. Accepted papers will be non-archival. The review process will be double-blind, with 2-3 reviewers per submission. OpenReview will be used for submissions; the workshop submission link and style-file details will be announced. See the Call for Papers for details.

Proposal Competitions

We invite two-page proposals for two competitions:

Iterative Improvement of AI Scientists: Design a framework through which an AI scientist learns to improve at a meaningful scientific task using existing data and knowledge as feedback, without requiring new experiments.
AI Credit Assignment in Science: Propose ways to improve citation, attribution, novelty assessment, review, and guardrails for AI-assisted science.

The top two proposals in each competition will receive podium presentations. Sponsors and cash award amounts are TBA. Read the competition requirements.

Invited Talks

The program includes six invited talks, each with 30 minutes for the talk and Q&A. Our speakers span materials science, physics, cosmology, and biology. Additional speakers will be announced.

Venkat Viswanathan

University of Michigan

AI × Materials

Anima Anandkumar

Caltech

AI × Physics

Risa Wechsler

Stanford University

AI × Cosmology

James Zou

Stanford University

AI × Biology

Panel

The panel topic, moderator, and panelists are TBA.

Program

The one-day program includes six invited talks, six contributed talks, four competition proposal highlights, one panel discussion, and two poster sessions. See the program overview for session formats and durations.

Community and Accessibility

We encourage in-person participation and will facilitate participation for virtual attendees and authors unable to travel. Accepted papers will be listed on this website, and presentation slides and talks will be shared as they become available, with speakers’ consent. Livestream and remote participation details will be announced.

We welcome researchers from varied scientific disciplines, institutions, career stages, and backgrounds. The workshop aims to connect machine learning researchers with the broader scientific community, including scientists attending an AI conference for the first time.

Post-Workshop Networking Event

We will host an evening networking event to continue conversations between scientists and AI researchers. Partners, venue, timing, and registration details will be announced.

Workshop Series

This edition builds on AI Scientists - Tools, Co-authors, or Founders? (ICML 2026) and Verification in the Age of AI Scientists (NeurIPS 2026). Together, these themes trace a progression from defining AI scientists, to verifying their outputs, to designing the workbenches through which they conduct science. Explore the AI for Science workshop series for previous editions.

Please follow us on X and LinkedIn for the latest news, or join our Slack community for discussions.