AI scientists now both produce and consume research — but the ecosystem they work in was built over three centuries for human readers. The limit on science is no longer model intelligence; it's that human-era ecosystem. A first-principles workshop on rebuilding how research is produced, represented, verified, composed, and evaluated — and why.
Co-developed CRISPR-Cas9 genome editing (Science 2013); now builds AI co-scientists for the wet lab — CRISPR-GPT and the LabOS AI-XR co-scientist — where a research artifact must be executable and verifiable enough to drive real physical experiments safely.
Leads UT Austin's Autonomous Mobile Robotics Lab and ran the AI-assisted peer-review experiments at AAAI and NeurIPS — a live test of review and reproduction at AI-scientist throughput, with human judgment kept in the loop.
Conducts foundational research on the algorithms and integrity of scientific peer review. His methods have been deployed across 200+ venues — including OpenReview — in the evaluation of over 100,000 papers.
Builds AI agents that reason, learn, and collaborate in complex environments — multi-agent collaboration, self-evolving agents that learn during deployment, and reasoning systems for scientific discovery. PhD from University of Washington with Yejin Choi.
Creator of TAOBench (deployed at Meta, PlanetScale, TiDB) and co-lead of AI-Driven Research for Systems (ADRS) — a live test of whether AI scientists can drive real engineering research, and how to measure them when they do.
Uses interactive theorem provers and dependent types to formally verify real-world programs, bringing proof-assistant rigor to AI-scientist claims.
Leads autonomous, data-driven scientific discovery at Ai2 (DataVoyager, AutoDS) and builds the AstaBench scientific-agent benchmark.
Leads Agent Research at Scale AI on LLM-agent evaluation and benchmarking at scale.
The modern research ecosystem — journals, peer review, conferences, citation, tenure tracks, grant cycles, the PDF itself — was built between the 17th and late 20th centuries for a single kind of participant: the human researcher. Each layer assumes a human author choosing what to write, a human reviewer choosing what to trust, and a human reader choosing what to build on. arXiv, OpenReview, and GitHub changed how those artifacts move, but not who they were written for, or by.
That assumption is no longer load-bearing. In 2025–2026, autonomous agents became a major contributor to research itself — producing papers, reviewing them, indexing them, and building on them. Agents now sit on both sides of every layer. The field is racing to make the model a better scientist while leaving the human-era ecosystem untouched — but the ecosystem, not raw model intelligence, is now the limiting layer, and the harder question is what ecosystem an AI scientist needs to do science in.
This workshop argues a simple thesis: if AI scientists are now a major contributor to research, every layer of the ecosystem — how research is produced, represented, verified, composed, and evaluated — has to be redesigned to treat AI scientists as first-class citizens, not retrofitted users of human-era institutions. Crucially, rebuilding each layer is a source of concrete ML research — benchmarks, datasets, evaluation protocols, and new agent methods — not institutional debate.
We organize the call around how a unit of research is produced, represented, verified, composed with other work, and evaluated at the level of the whole network. One foundational track measures the producer; the rest rebuild the layers its output flows through. We prioritize contributed empirical and technical work; position and vision papers are welcome but capped as a minority track.
How AI scientists actually behave today — and whether they understand the science or only reproduce its narrative.
The right unit of research when the consumer is an agent — at once executable and verifiable.
Verification, reproduction, and peer review at agent throughput — with human judgment kept in the loop.
Turning isolated AI scientists into a collaborating network — contributions that compose, connected by typed dependencies rather than citation-as-string.
Measuring and steering the multi-actor network — ecosystem-level benchmarks, contribution attribution, and incentive-compatible mechanisms so truthful review and reusable artifacts become the dominant strategy.
Research papers (primary track). 9-page long or 4-page short papers in ICLR/NeurIPS format; references and appendices excluded from the page count. New methods, benchmarks, datasets, systems, and evaluation protocols. This is the substance of the program.
Position papers (minority track). 4 pages. First-principles arguments about what the ecosystem should become — welcome and spotlighted, but capped so the workshop platforms contributed research rather than becoming a discussion-only venue.
Review & policy. All submissions handled via OpenReview, double-blind, ≥3 reviewers per paper. Non-archival; dual submission allowed per ICLR/NeurIPS workshop norms. The LLM policy follows the NeurIPS 2026 main track: AI-assisted writing is permitted with disclosure, and AI-authored submissions must designate a human corresponding author who certifies content.
Submission portal opens 2026-08-15. Link to be posted here.
Dates are tentative and follow NeurIPS 2026 workshop deadlines. All times Anywhere on Earth (AoE).
Six 40-minute invited talks, a structured debate mid-morning, a moderated panel late in the day, and a closing two-hour poster & demo session for contributed papers.
| 09:00 – 09:10 | Opening remarks |
| 09:10 – 10:30 | Invited Talks 1–2 (40 min each) |
| 10:30 – 10:50 | Coffee break |
| 10:50 – 11:30 | Debate — a provocative motion on the future of the AI-driven research ecosystem (two sides + audience vote) |
| 11:30 – 12:50 | Invited Talks 3–4 |
| 12:50 – 13:50 | Lunch |
| 13:50 – 15:10 | Invited Talks 5–6 |
| 15:10 – 15:20 | Break |
| 15:20 – 16:00 | Panel + moderated discussion |
| 16:00 – 16:05 | Closing remarks |
| 16:05 – 18:00 | Poster & Demo Session (contributed papers) |
