Skip to main content

1st Workshop on Scientific Superintelligence

Evidence, Expert Baselines, and Recursive Improvement

ICLR 2027 Workshop Proposal

When does an AI discovery system outperform expert scientists, and when does its next version become a better discoverer?

April 29–30, 2027 · San Francisco, California (exact day and room to be announced)

News

  • [Oct 2026] Invited Speakers Confirmed: Wei Zhou, Guannan Zhang, Haotian Zhang, Jian Tang, and Shuchao Bi have confirmed attendance.
  • [Oct 2026] Workshop Proposal Submitted: The 1st Workshop on Scientific Superintelligence has been proposed to ICLR 2027. Acceptance decisions are expected on November 29, 2026.
  • [Oct 2026] Website Launched: Welcome to the website of the proposed SSI@ICLR2027 workshop.

About

When does an AI discovery system outperform expert scientists, and when does its next version become a better discoverer? The 1st Workshop on Scientific Superintelligence will turn these ambitions into two testable questions: comparative capability at matched resources and recursive improvement under fixed external evaluation. Scientific superintelligence is the hypothesis being examined: superior performance must be tied to explicit tasks, credible expert-team and algorithmic baselines, and independently checked outcomes.

AI4Sci, AI4Mat, AI4Chem, AI4Math, AI4Physics, AI4AI, and self-driving laboratories provide test settings for these questions. Comparisons will account for compute, laboratory access, elapsed time, and human effort, and distinguish scientific novelty, task completion, and transferable capability. Studies of individual representations, feedback-learning methods, or evaluators are welcome when they inform either question.

Five invited talks, contributed presentations, posters, an evidence clinic, two focused panels, and working groups will produce one primary artifact: a reporting template for comparative capability claims with three worked examples. Physical experiments, formal proof, and computational/AI research will share comparison fields while retaining their own verification standards. The workshop will sharpen protocols and expose failure modes that researchers can study after the meeting.

Two Pillars: Comparative Capability and Recursive Improvement

  1. Beyond-expert claims at matched resources. What task and comparator would make a claim of exceeding expert performance meaningful? We seek expert-team, human–AI-team, and algorithmic baselines with documented expertise and information access. Comparisons should specify compute, experiment capacity, elapsed and human time, stopping rules, and independent outcomes. Where resources cannot be exactly matched, cost–performance curves and explicit differences are preferable to an unqualified superiority claim. Temporal splits and provenance checks should separate new discovery from retrieval, leakage, and selective reporting.
  2. Recursive improvement under fixed external evaluation. Which changes to models, representations, memory, tools, or policies improve discovery beyond an earlier version and a static baseline? We welcome RL, active learning, generative modeling, optimization, and agents learning from scarce, noisy, delayed, and negative feedback. Studies should freeze external criteria, separate adaptation from held-out assessment, account for update costs, and check regression and transfer. The external evaluator may itself be imperfect: its uncertainty and failure modes should be assessed independently of the system it scores. Evaluator modification or higher self-scores alone do not establish better discovery.

Call for Papers

This call is provisional and takes effect if ICLR 2027 accepts the workshop (decisions expected November 29, 2026).

We invite submissions on the workshop’s two testable questions: comparative capability at matched resources and recursive improvement under fixed external evaluation. The call will prioritize ongoing or novel research, including methods, theory, new experiments, replication, negative results, and well-supported perspectives. We welcome contributions from representation and generative-model researchers; RL, active-learning, and agent researchers; reliability, formal-methods, and evaluation researchers; domain scientists and laboratory practitioners; optimization and theory researchers; and junior and under-resourced researchers. Participants can contribute to one methodological component and gain feedback from complementary communities; a complete autonomous laboratory or end-to-end discovery system is not required.

Key Dates (tentative)

  • CFP Launch and OpenReview Setup: December 7, 2026
  • Submission Deadline (both tracks): February 1, 2027
  • Decisions, Author Notifications, and Collection of Final Public Versions: February 22, 2027
  • Accepted Papers Public on OpenReview: February 25, 2027
  • Workshop Date: April 29 or 30, 2027 (conference assignment pending)

All deadlines are 23:59 Anywhere on Earth (AoE).

Submission Site

Papers will be submitted through OpenReview. The submission link will be posted here when the portal opens.

Scope

Scope criterion. Submissions should inform either comparative capability at matched resources or recursive improvement under fixed external evaluation. Contributions on representations, scientific-feedback learning, and evaluators are welcome when they advance one of these questions, with evidence appropriate to their stated scope. Individual components are welcome; short papers need not implement an end-to-end system.

Scientific domains as evaluation environments

The seven scientific areas (AI4Sci, AI4Mat, AI4Chem, AI4Math, AI4Physics, AI4AI, and self-driving laboratories) are shared test settings rather than separate tracks. They are organized into three validation environments:

Physical Experiments

Representative tasks: Materials and catalyst selection; synthesis; device fabrication; laboratory planning.

Independent evidence: Controls, repeat preparation, calibration, uncertainty, and independent measurement or replication.

Formal Proof

Representative tasks: Conjecture evaluation, formalization, proof search, and verified algorithmic claims.

Independent evidence: Precise statements and assumptions, trusted proof checking, and inspection of statement fidelity.

Computational/AI Research

Representative tasks: Equation discovery, scientific simulators, architecture or optimizer design, and research-tool updates.

Independent evidence: Locked tests, held-out tasks, reproducible execution, ablations, and matched tuning and compute.

Primary output and advances sought

The reporting template will require task and claim, expert and algorithmic baselines, available information, resource envelope, human intervention, failure history, update mechanism, fixed external checks, uncertainty, and transfer scope. Three worked cases will identify what each comparison supports and what decisive evidence remains missing. We will annotate autonomy using established self-driving-laboratory and AGI frameworks and the autonomy spectrum discussed at ICML 2026’s AI for Science workshop, connecting the claimed level to required evidence rather than creating another autonomy vocabulary. Short contributions may advance one field or component; full autonomous workflows and claims of expert superiority are not prerequisites for participation.

Submission Guidelines

Format:

All submissions must be a single PDF file. The workshop has two non-archival tracks, submitted through OpenReview. The main text must be self-contained; reviewers are not required to read beyond it.

  • Research Track: Up to eight pages of main text, with references and optional appendices excluded. Appropriate contributions include discovery algorithms, scientific representations, learning theory, evaluation methods, and experimentally grounded systems.
  • Tiny/Short Paper Track: Two to four pages of main text, with references excluded. We welcome a modest but self-contained result, a simple implemented idea, an additional experiment or re-analysis, a failed replication, or a fresh evidence-based perspective. This track shares the February deadline; it does not introduce a later acceptance round.
ICLR 2027 Guidelines

In line with the ICLR 2027 workshop guidelines, our tiny/short paper track prioritizes “less-than-full-conference” contributions and aims to broaden participation, especially from early-career researchers and authors outside the usual ML publication circuit. AI assistance is permitted, but tiny/short papers must be primarily human-authored and cannot be AI-generated. Authors of accepted papers may apply for ICLR 2027 Financial Assistance through a separate application; details will be posted on iclr.cc/Conferences/2027.

Style file:

Both tracks will use the official ICLR 2027 LaTeX template and anonymous review. Include references and supplementary material in the same PDF. Please change the template’s running header, which refers to the main conference, to refer to the workshop (e.g., “Under review at the SSI Workshop at ICLR 2027”). Submissions that alter the ICLR style (e.g., by decreasing margins or font sizes) or exceed the page limits may be rejected without review.

Review process:

Every submission will receive at least two human reviews, with a third review or independent adjudication where needed. Reviewers will assess scientific relevance, soundness, clarity of claims, adequacy of validation, novelty relative to disclosed prior work, and value for discussion. Short papers will be judged against their stated scope. Final selection will depend on quality and discussion value.

Evidence expectations:

Wet-lab access and large compute budgets will not be prerequisites: a careful analysis, formal result, or openly reproducible small-scale study can make a strong contribution. For empirical discovery claims we will request appropriate baselines, uncertainty, provenance, resource costs, and the extent of human intervention. Artifacts will be encouraged where release is feasible. Domain experts will assess scientific validity separately from claims made by an agent or its internal scoring model. We will encourage release of unsuccessful trials and limitations to make selective reporting visible. Work involving experiments should describe relevant oversight and operating constraints. Discussion of autonomy will include human accountability, misuse risks, laboratory safety, and the ability to audit or interrupt a system, proportionate to the work presented.

Prior work and non-archival policy:

Previously published work may provide context, but a submission must clearly identify new content; the workshop will not primarily re-present papers from the ICLR main track or other ML conferences. Acceptance will not constitute an archival proceedings publication. Accepted papers will remain publicly accessible on OpenReview and be listed on the workshop website. Authors must check the policies of any concurrent venue. We will publish session and talk titles before the conference program is finalized so attendees can choose sessions by content.

Visibility:

Submissions and reviews will not be public. Only accepted papers will be made public on OpenReview.

Double-blind reviewing:

All submissions must be anonymized and may not contain any identifying information that violates the double-blind reviewing policy. This applies to supplementary and linked material as well, including code; if you link to external material, you are responsible for ensuring anonymous browsing. Please do not include acknowledgements at submission time, and cite your own prior work in the third person. Papers that violate this policy may be rejected without review.

Conflicts of interest:

An organizer will not handle a submission from their own institution or organization, including any part of the same corporation. Coauthorship, advising, collaboration, personal, and financial conflicts will also require recusal. A non-conflicted chair will assign reviewers and make the decision. If all organizers are conflicted, an independent external decision chair will handle the paper. Organizer-authored submissions will receive external handling, and organizers will not present talks at this workshop.

AI participation policy:

The workshop studies AI conducting research, including AI-generated hypotheses, code, proof candidates, and experimental plans. Submissions in both tracks require accountable human authors. AI will not serve as a primary author, peer reviewer, or decision maker. Tiny/short papers must be primarily human-authored and cannot be AI-generated; AI assistance is permitted. All authors will disclose substantive AI use and how outputs were checked, following the ICLR 2027 author guidance.

Reviewers must write their own substantive assessments. Any permitted AI assistance in editing reviews must follow the ICLR disclosure requirements, including preservation of the original assessment and model interactions. Confidential submissions may not be entered into external AI services without appropriate permission and confidentiality protections. Research on AI-assisted review can be presented as research, but will not replace the workshop’s human review process.

Presentation:

ICLR 2027 workshops are held in person. All accepted contributions receive a poster slot; six receive orals and ten receive lightning slots. Selections will cover domains, methods, and career stages, with at least two lightning slots reserved for tiny/short papers, subject to suitable submissions. For exceptional visa or other constraints, we will offer a poster proxy and a short prerecorded presentation on the website.

Evidence clinic:

Before the workshop, we will select three volunteered cases: a physical experiment, a formal proof, and a computational or AI research experiment. Each author will prepare a task, expert or static-system baseline, update history, resource envelope, independent check, and unresolved failure mode. For each 15-minute case, five minutes of presentation will be followed by ten minutes of audience discussion with a method researcher and a domain or validation expert. Cases will use the same core reporting fields while keeping environment-specific checks. No live laboratory experiment is required.

Online access:

Accepted papers, optional posters, captioned prerecorded videos, and discussion notes will be available online. An asynchronous question form will support contributions from those unable to attend. We will request permission and conference support to record invited talks; if recording is unavailable, we will publish consenting speakers’ slides and written summaries. The access plan does not depend on live streaming or a hybrid event.

Outcomes:

The primary outcome will be a public reporting template for comparative capability claims with three complete worked examples covering physical experiments, formal proof, and computational/AI research. A starter template will be prepared before the meeting, revised in the clinic and working groups, and published within four weeks. The accompanying report will capture disagreements, validation boundaries, and possible follow-up benchmark projects. Contributors will be credited and may opt into further collaboration. The organizing team will maintain the materials for at least one year. An online community meeting within three months will assess uptake and replication opportunities.

Contact:

For any questions, please contact the primary contact, Ricky Renjie Li, at renjie2@illinois.edu.

Accepted Papers

(click to expand)

Accepted papers will be listed here after the review period (by February 25, 2027).

Program Committee

(click to expand)

The program committee will be announced soon.

Interested in reviewing for SSI@ICLR2027? We especially welcome reviewers from the natural sciences and mathematics. Please email us at renjie2@illinois.edu.

Schedule

To Be Announced

The detailed schedule will be posted closer to the workshop. The one-day, single-track program combines invited talks with contributed research, an evidence clinic, and structured discussion: five invited talks, two contributed oral sessions, a poster lightning session including tiny/short papers, two poster sessions, two focused panels, and cross-domain working groups.

Invited Speakers

The invited program brings together academic, national-laboratory, and industry perspectives on scientific AI. Talk titles will be announced soon.

Wei Zhou

Associate Professor
Virginia Tech

Guannan Zhang

Distinguished Staff Scientist
Oak Ridge National Laboratory

Haotian Zhang

Senior Research Scientist
NVIDIA Cosmos Lab

Jian Tang

Associate Professor
HEC Montréal and Mila

Shuchao Bi

Research Scientist
Meta Superintelligence Labs

Panelists

The two panels will address When does AI outperform expert scientists? and How can we establish that an updated system discovers better? The proposed panelist roster is:

Hongteng Xu

Renmin University of China

Chen Cai

Google

Xuewei Wu

Meta

Ivory Zhang

University of British Columbia

Yuanhao Zhou

Carnegie Mellon University

Zhidi Lin

The Education University of Hong Kong

Yuxing Wang

NVIDIA

Yanan Li

Zhejiang Lab

Hao-Chung Kuo

National Yang Ming Chiao Tung University

Ceyao Zhang

Peking University

Zhen Li

The Chinese University of Hong Kong, Shenzhen

Workshop Organizers

Hanghang Tong

University of Illinois Urbana-Champaign

Ricky Renjie Li

University of Illinois Urbana-Champaign

Tianshu Yu

The Chinese University of Hong Kong, Shenzhen

Jiawei Zhang

University of Wisconsin–Madison

Huan Zhang

University of Illinois Urbana-Champaign

Chudi Zhong

University of North Carolina at Chapel Hill

Shangding Gu

Shanghai Jiao Tong University

Ge Liu

University of Illinois Urbana-Champaign