The Future of the AI-Driven Research Ecosystem

Why this workshop

AI Scientist v2 pushed a fully AI-generated paper through workshop peer review. Biomni runs biomedical workflows end to end. The Virtual Lab assembled a team of AI agents that designed nanobodies later validated at the bench. Yet every one of these systems publishes into the same ecosystem built for human bandwidth three centuries ago: the PDF as the unit of publication, citations as the only link between works, review cycles measured in months, archives and conferences built around human reading.

Science compounds through many minds and successive generations inheriting, reusing, and iterating on each other's work, and that compounding runs on the shared ecosystem underneath. The field is racing to make the individual AI scientist smarter while leaving that human-era ecosystem untouched — so the ecosystem itself now throttles the speed of science. This workshop takes the ecosystem, rather than the agent, as the unit of redesign.

Why now

AI-scientist systems and the benchmarks tracking them are launching this year, each shipping its own ad hoc conventions for artifacts, traces, and verification. Once two or three frontier labs converge on vendor-specific defaults, switching costs make redesign infeasible — that is how the arXiv PDF and citation-as-string became permanent fixtures.

The window for setting first-principles ecosystem standards is the next 12–24 months.

And NeurIPS is where most of these systems are designed: if the ecosystem question is not posed here, it does not get posed in one place at all.

Five questions, one path

We organize the call around the path a unit of research takes through the ecosystem: how it is produced, represented, verified, composed with other work, and evaluated at the level of the whole network.

• Produce. Measure how AI scientists actually behave before redesigning anything around them. The open question is not whether an AI scientist can produce a paper but whether it understands the science — forming transferable abstractions and capturing causal structure rather than reproducing narrative from surface statistics.
• Represent. Design representations native to the AI scientist: research as typed, structured objects — claims, evidence, methods, dead ends — that are executable and verifiable at once, composable with other work, and queryable at the claim level rather than the paper level.
• Verify. Review and reproduction at AI-scientist throughput, with hybrid human–AI adjudication that automates structural checks and reserves human judgment for taste.
• Compose. Turn isolated agents, human and AI alike, into a collaborating network whose contributions connect through typed dependencies instead of citation-as-string.
• Evaluate. Measure the network itself: how fast knowledge compounds across many actors, and whether credit flows to the contributions that caused progress.

The program

The lineup is compact by design — six invited talks, one debate, one panel — with verification and peer review as the workshop's center of gravity. Confirmed speakers include Le Cong (Stanford), who co-developed CRISPR-Cas9 and now builds AI co-scientists for the wet lab; Nihar Shah (CMU), whose peer-review methods have evaluated 100,000+ papers across 200+ venues; Joydeep Biswas (UT Austin), who ran the AI-assisted peer-review experiments at AAAI and NeurIPS; Lianhui Qin (UC San Diego) on multi-agent collaboration; Audrey Cheng (UC Berkeley) on AI-driven research for systems; and Yao Li (Portland State) on making AI-scientist claims machine-checkable. Yue Zhang (Scale AI) and Bodhisattwa Majumder (Ai2) close the day on a moderated panel.

A structured debate mid-morning sets up the day's motion, and a two-hour poster & demo session closes it — contributed papers and live demos side by side.

Join us

The call welcomes 2-page Tiny Papers and 4-page position papers, alongside non-traditional contributions: datasets, negative results, and demos. Submissions run through OpenReview — opening August 15, 2026 and closing September 15, 2026 (AoE; dates tentative, following NeurIPS workshop deadlines) — and travel awards prioritize students and under-resourced institutions.