A proposed ICLR 2027 workshop
Autonomous research and recursive self-improvement across AI4AI and AI4Science — grounded in verification, evaluation integrity, and safety.
00 · News
01 · About
Research agents are beginning to carry out increasingly complete loops: reading the literature, forming hypotheses, designing and running experiments, interpreting results, and communicating findings. Those loops now appear both in AI4AI — improving models, data, algorithms, and computing systems — and in AI4Science, from mathematics and materials to biology, chemistry, and physics.
Autonomy alone is not recursive self-improvement. The stronger claim begins when a system's discoveries improve the models, tools, evaluators, or research strategies used in its own future cycles, and when those gains persist rather than disappear under new tasks or stronger controls. AI4Research puts both questions in one program: how much of research can be made autonomous, and when does that autonomy produce cumulative improvement?
Our intellectual spine is verification. AI4AI often offers executable feedback such as unit tests, benchmark scores, and wall-clock measurements; AI4Science adds simulators, theorem checkers, instruments, and experimental evidence. Across both domains, evaluation integrity must remain a first-class research topic: results need provenance, contamination controls, human-effort accounting, reproducibility, and defenses against reward hacking.
AI4Research connects communities that are often discussed separately. Autonomous-research work asks how agents conduct discovery; recursive-self-improvement work asks whether improvements compound. AI4AI and AI4Science provide complementary proving grounds for both. The workshop focuses on the shared methods, evidence standards, infrastructure, and safety boundaries that let progress transfer across all four.
From agents that execute a research loop to systems whose results measurably improve later loops.
From models and ML infrastructure to scientific reasoning, simulation, and experiment-grounded discovery.
02 · Topics
We invite submissions including, but not limited to, the following areas.
Systems that read literature, form hypotheses, design and run experiments, interpret evidence, and communicate findings; architectures, workflows, and failure analyses.
Agents for data, model design, training and post-training, debugging, AutoML, GPU kernels, compilers, serving systems, and the broader ML stack.
Research agents for mathematics, physics, chemistry, biology, medicine, materials, and climate; simulation, theorem proving, lab workflows, and experiment-grounded discovery.
Self-referential agents, automated system design, open-ended evolution, and empirical tests of whether improvements persist, transfer, compound, or plateau.
Tasks, environments, and metrics for ideation, experimentation, replication, engineering, and scientific discovery; contamination-resistant and long-horizon design.
Specification gaming, benchmark exploits, evaluator attacks, and hardened harnesses, co-evolving evaluators, execution-based verification, and evidence provenance.
Prompt and program optimization, workflow and topology search, tool use, memory, test-time-compute scaffolds, and self-modifying research harnesses.
Evaluation of research artifacts, LLM-as-judge, automated reviewing and meta-reviewing, replication agents, and evaluating the evaluators.
Autonomy levels, time-horizon and uplift studies, human-comparison methodology, compute and human-effort accounting, and economics of AI-assisted research.
Risk thresholds, oversight of self-improving systems, dual-use scientific capabilities, secure experiment execution, and policy for AI-authored research.
Division of labor, trust calibration, research taste, attribution and authorship, accessibility, and keeping humans meaningfully in the loop.
Open-source agent frameworks, compute and laboratory environments, shared run archives, interoperable tools, and reproducibility standards for agent-conducted research.
03 · Call for Papers
The CFP below is tentative and becomes official upon workshop acceptance by ICLR 2027.
04 · Important Dates
All deadlines are 11:59 pm Anywhere on Earth (AOE) and tentative until the ICLR 2027 workshop decisions are announced.
05 · Schedule
| Time | Session | Duration | Details |
|---|---|---|---|
| 9:00 – 9:10 | Opening remarks | 10 min | Organizers |
| 9:10 – 9:50 | Keynote 1 — End-to-end research agents | 40 min | Invited talk + Q&A |
| 9:50 – 10:35 | Contributed talks I | 45 min | 3 × 15 min oral presentations |
| 10:35 – 11:05 | Coffee break & posters | 30 min | Poster session I |
| 11:05 – 11:45 | Keynote 2 — AI4AI: models, systems & infrastructure | 40 min | Invited talk + Q&A |
| 11:45 – 12:15 | Tiny-paper lightning talks | 30 min | 6 × 5 min — the on-ramp track, on stage |
| 12:15 – 13:30 | Lunch break | 75 min | — |
| 13:30 – 14:10 | Keynote 3 — AI4Science: verified scientific discovery | 40 min | Invited talk + Q&A |
| 14:10 – 15:00 | Safety session — risks of recursive self-improvement | 50 min | Invited talk + moderated discussion: RSI risk thresholds, evaluation integrity under self-modification, dual-use discovery, and oversight |
| 15:00 – 15:45 | Poster session II & coffee | 45 min | All accepted papers |
| 15:45 – 16:30 | Contributed talks II | 45 min | 3 × 15 min oral presentations |
| 16:30 – 17:20 | Panel — “Who is doing the research?” | 50 min | Panelists + moderator, audience Q&A |
| 17:20 – 17:30 | Closing remarks & awards | 10 min | Best Paper and Best Tiny Paper announcements |
06 · Speakers
Yilun Du and Zifan Wang are confirmed. The remaining names are prospective invitees, clearly labeled below. The final program features three keynotes, one safety-session speaker, and four to five panelists drawn from the confirmed roster. No organizer will speak at the workshop.
07 · Panel
A 50-minute moderated discussion with four to five panelists drawn from the confirmed speaker roster, balancing capabilities, systems, evaluation, and safety perspectives. Moderated by an organizer. Candidate questions:
08 · Organizers
The organizing team is being assembled ahead of the ICLR 2027 proposal deadline, targeting diversity of affiliation, seniority, geography, and viewpoint. Profiles being recruited:
Interested in co-organizing or reviewing? Get in touch via the contact address below.
09 · FAQ