End-to-end autonomous discovery
Architectures closing the loop from literature and data through hypothesis, experiment, analysis and write-up — including honest post-mortems of systems that discovered nothing.
Cross-disciplineSystems
Mission log · 01
News
-
Workshop site is live. The FAR proposal is in preparation for the ICLR 2027 Call for Workshops, which is expected around September 2026. Nothing on this page implies acceptance.
-
Draft topics of interest, tentative programme and panel questions published for community feedback. Suggestions for topics, speakers and shared evaluation artifacts are welcome at the contact address below.
-
Expected Dec 2026
ICLR 2027 workshop decisions. If FAR is accepted, the full call for papers, the OpenReview submission link and confirmed dates go live here.
Mission · 02
About the workshop
Between 2023 and 2026, AI systems moved from assisting scientists to producing candidate scientific content at scale. A general-purpose reasoning model produced the core construction disproving the Erdős unit-distance conjecture, open since 1946 (OpenAI, May 2026), a result Timothy Gowers called "a milestone in AI mathematics." Two multi-agent co-scientists published wet-lab-validated results in Nature in the same month: Google's AI co-scientist, which independently recovered an unpublished antimicrobial-resistance mechanism, and FutureHouse's Robin, which autonomously generated the hypotheses, analyses and figures identifying ripasudil as a dry-AMD candidate while humans held the pipette.
In the same window the other half of the picture came into focus. Frontier models score 77% on olympiad-style science questions and 25% on open-ended research tasks (OpenAI's FrontierScience, December 2025). LLM-generated research ideas are rated more novel than human ideas before execution and worse on every metric after it. Fabricated citations in the published literature rose roughly tenfold in three years. Agent-written but unsound papers pass automated review. Self-driving laboratories still sit at autonomy Level 2–3 of 5, and as of early 2026 no AI-originated material had entered commercial production.
Google DeepMind named the structure of the problem in a July 2026 policy essay — an argument this workshop takes as something to test, not a peer-reviewed finding it inherits: AI agents are conjecture machines, and as hypothesis generation approaches free, refutation — "the slow, costly business of testing whether an idea survives contact with reality" — becomes the binding constraint on scientific throughput. That asymmetry is the spine of this workshop. Mathematics automated first because verification is formal and nearly free; the open question is what the equivalent of a proof kernel looks like in biology, chemistry and physics.
FAR convenes three communities that rarely share a room: people building end-to-end discovery systems, domain scientists across mathematics, the physical sciences, chemistry, materials and medicine, and the benchmark, verification and epistemics researchers who must certify what those systems produce. Our three goals: (1) compare autonomous-discovery architectures on evidence rather than demos; (2) advance evaluation and verification for discovery-shaped, long-horizon claims; (3) treat safety and epistemics — fabricated science, dual use, validating AI-generated claims — as a first-class session rather than a closing caveat.
Survey area · 03
01
02
03
04
05
06
07
08
09
10
11
12
Topics of interest
Twelve areas, spanning the disciplines. We especially encourage validated discoveries with an explicit account of what the system did versus what humans did, rigorous negative results, position papers on evaluation and epistemics, and work from domain scientists outside the machine-learning community.
AI co-scientists & research teams
Tournaments, debate, role-specialised pipelines and PI/team hierarchies; long-horizon memory and world models; empirical studies of how working scientists actually collaborate with these systems.
AgentsHuman–AIAutomated mathematical discovery
Formal and informal theorem proving, autoformalisation at document and library scale, conjecture generation, and evidence about which open problems are and are not within reach.
MathematicsFormal methodsDiscovery in the physical sciences
Agents for physics, astronomy, earth and climate science; symbolic regression and equation discovery; large simulation campaigns; survey-scale observational triage at Rubin/LSST data rates.
PhysicsAstronomyChemistry & materials agents
Retrosynthesis and synthesis planning, generative inverse design, agentic control of computational experimentation, and the gap between predicted and realised, useful materials.
ChemistryMaterialsBiology & medicine
Protein and genome design agents, target identification and drug repurposing, virtual-cell and in-silico experimentation loops, agent-driven analysis of large biological datasets.
BiologyMedicineHypothesis generation & research taste
Generating, ranking and pruning hypotheses; measuring novelty, feasibility and expected value; the ideation–execution gap; whether research taste is learnable or merely nameable.
MethodsCognitionExperiment design under uncertainty
Active learning and Bayesian experimental design for agents; budget- and risk-aware experiment selection; orchestration across computational, simulated and physical experiments.
MethodsPlanningLiterature-scale reasoning
Retrieval, synthesis, contradiction and replication-failure detection across millions of papers; grounding and attribution; turning the published record into an actionable tool library.
Cross-disciplineRetrievalBenchmarks, evaluation & reproducibility
Evaluation for long-horizon, open-ended, discovery-shaped tasks; contamination and saturation; cost-normalised comparison; reproducibility of agent results as a first-class measurement.
EvaluationBenchmarksEpistemics, verification & trust
Claim-level auditability, mechanism fidelity versus outcome correctness, AI-assisted peer review, provenance standards, and the effect of agent-generated volume on the scholarly record.
EpistemicsVerificationSafety, dual use & governance
Biological, chemical and cyber misuse risk from capable scientific agents; structured access and gated release; evaluations for agentic uplift; policy and funding implications of accelerated discovery.
SafetyPolicyAutomation of machine-learning research is welcome as a case study in a discipline where verification is cheap, positioned comparatively against disciplines where it is not — but it is not the workshop's centre of gravity.
Expedition brief · 04
Call for papers
Submissions open upon workshop acceptance (expected December 2026). The details below are the planned call and are subject to the ICLR 2027 workshop guidelines once published. No submission link exists yet; this page will be updated the day it does.
Research Track
Up to 8 pages excluding references and appendix, ICLR 2027 style. Original systems, methods, benchmarks, empirical studies and negative results.
Short / Position Track
Up to 4 pages excluding references. Position papers, early results, evaluation critiques, benchmark and dataset proposals. Per ICLR policy, papers in this short format must be primarily human-authored.
Review process
Double-blind on OpenReview. Each submission receives at least three reviews from a program committee of roughly 40–60 reviewers, with domain-matched assignment: a chemistry submission is read by at least one chemist. Organisers do not handle submissions from their own institutions.
Archival status
Non-archival. Accepted papers are posted on OpenReview with author consent and may be submitted elsewhere afterwards. Work already published at ICLR or another major ML conference is out of scope; concurrent submission elsewhere is permitted and must be disclosed.
Review criteria include validation
Beyond the usual criteria, reviewers assess: does the paper state what the system did versus what humans did; are discovery claims supported by verification independent of the generating system; are costs and failure rates reported.
Awards
Best paper, best negative result, and best evaluation artifact (benchmark, harness or reproducibility study). Certificate-only unless sponsorship is secured — we will not advertise prize money we do not have.
AI-assisted submission policy
AI assistance in writing, coding, analysis and ideation is permitted and expected in this field, and must be disclosed in a required one-paragraph AI involvement statement that does not count toward the page limit. The statement must specify (i) which AI systems were used, with versions; (ii) the nature of the AI contribution, mapped to the research stages it touched — literature, ideation, experimental design, execution, analysis, writing; and (iii) the human oversight applied, including who verified each empirical claim and how. Papers reporting results produced substantially by an autonomous system must also state an autonomy level and, where possible, supply execution transcripts sufficient for a reviewer to check the chain from data to claim. Fabricated citations or unverified generated results are grounds for desk rejection; we plan to spot-check references on accepted papers and report aggregate findings, never individual authors. Submissions touching biological, chemical or cyber capability uplift are flagged by authors at submission and screened against a published checklist.
Trajectory · 05
Important dates
All deadlines are tentative and end-of-day AOE. They will be fixed against the ICLR 2027 global workshop calendar if the workshop is accepted.
- ICLR 2027 workshop decisions December 2026 (tentative) FAR is a proposal until this point. Everything below is conditional on acceptance.
- Submissions open on OpenReview December 2026 (tentative) Link published here as soon as the venue exists. No dead buttons before then.
- Paper submission deadline February 3, 2027 (AOE) — tentative Both tracks. Any extension will be shown here with the original date struck through, not silently edited.
- Notification of acceptance March 1, 2027 (AOE) — tentative Bound by the ICLR global workshop notification deadline, which is non-extendable.
- Camera-ready deadline April 1, 2027 (AOE) — tentative Including the required AI involvement statement.
- Workshop day April 2027 — tentative One day, in person, with ICLR 2027. Location to be announced by ICLR.
Day plan · 06
Tentative schedule
| Time | Session | Duration | Details |
|---|---|---|---|
| 09:00 – 09:10 | Opening remarks | 10 min | Organisers: the generation–validation gap, and what this day is for |
| 09:10 – 09:50 | Keynote 1 — Discovery in the formal sciences | 40 min | Invited talk + Q&A |
| 09:50 – 10:30 | Keynote 2 — AI co-scientists in biology and medicine | 40 min | Invited talk + Q&A |
| 10:30 – 11:00 | Coffee break & Poster session I | 30 min | Spotlight posters, all disciplines |
| 11:00 – 12:00 | Contributed talks I | 60 min | 4 × 15 min orals, selected across disciplines |
| 12:00 – 13:30 | Lunch break | 90 min | Optional mentoring tables for junior authors |
| 13:30 – 14:10 | Keynote 3 — Chemistry, materials and the physical sciences | 40 min | Invited talk + Q&A |
| 14:10 – 14:55 | Contributed talks II | 45 min | 3 × 15 min orals, including a designated best-negative-result slot |
| 14:55 – 15:35 | Poster session II & coffee | 40 min | All accepted papers — poster space for up to approximately 80 posters across the two sessions |
| 15:35 – 16:25 | Safety & epistemics session | 50 min | Invited talk (20 min) + 3 lightning talks (5 min each) + 15 min moderated discussion: fabricated science, dual use, validating AI-generated claims |
| 16:25 – 17:25 | Closing panel — Generation is cheap. Who validates? | 60 min | Runs straight out of the safety session with no break; 5 panellists + organiser moderator; final 20 min for audience questions |
| 17:25 – 17:35 | Closing remarks & awards | 10 min | Best paper, best negative result, best evaluation artifact |
The final programme is posted after author notifications. The day is deliberately back-loaded toward validation: the safety & epistemics session runs directly into the closing panel, so the day ends on the question of what can be trusted rather than on capability demonstrations.
Crew · 07
Pushmeet Kohli
Google DeepMind
to be invited
John Jumper
Anthropic
to be invited
Sam Rodriques
FutureHouse · Edison Scientific
to be invited
Marinka Zitnik
Harvard Medical School
to be invited
Javier Gómez-Serrano
Brown University
to be invited
Connor Coley
MIT
to be invited
Miles Cranmer
University of Cambridge
to be invited
Carla Gomes
Cornell University
to be invited
Invited speakers
The list below records who we intend to invite. As of August 2026 no one has been approached and no one has agreed to participate; every name is marked to be invited, and this page will only ever mark a speaker confirmed with their explicit agreement. Affiliations will be re-verified at invitation time — several changed during 2026. Organisers do not give talks at their own workshop.
Rendezvous · 08
Panel — Generation is cheap. Who validates?
Sixty minutes, scheduled as the day's closing plenary and running directly out of the safety and epistemics session, so the discussion arrives with the evidence already on the table. Five panellists spanning mathematics, biology, chemistry and materials, and evaluation and safety, plus one designated sceptic of the workshop's own optimistic premises. Moderated by an organiser — ICLR rules permit organisers to moderate but not to present. The final twenty minutes are reserved for audience questions collected on cards during the day.
Panellists and reserves — all to be invited:
- Andrew White · FutureHouse
- Regina Barzilay · MIT
- Vivek Natarajan · Google DeepMind
- Ekin Dogus Cubuk · Periodic Labs
- Max Welling · U. Amsterdam · CuspAI
- Kyunghyun Cho · NYU
- Anima Anandkumar · Caltech · bench alternate
- Yolanda Gil · USC/ISI · bench alternate
- Terence Tao · UCLA · bench alternate
Questions on the table
- Where is the line between discovery and retrieval? Does the Erdős unit-distance result sit on a different side of it from Robin's ripasudil — given that a 2025 "ten Erdős problems solved" claim turned out to be literature retrieval?
- Mathematics has Lean. What is biology's kernel? Is verification-in-the-loop achievable in the empirical sciences, or does physical experiment cap how autonomous any biology agent can be?
- What lives in the 52-point gap between 77% on olympiad science and 25% on open-ended research — problem formulation, taste, tolerance for ambiguity, tool use, or something unnamed? Which part is closing fastest?
- Should the field stop optimising hypothesis generation? LLM ideas rate as more novel before execution and worse after it.
- Is the scholarly record already net-worse? Fabricated citations rose roughly tenfold in three years. What would have to be true in 2028 for the answer to flip?
- Is a workshop on automating science a workshop on accelerating a named risk? Automated R&D appears in frontier-lab governance frameworks alongside bioweapons and loss of control. What does that oblige of the authors in this room?
Crew · 09
Organizers
Yuxuan Zhang
PhD student, University of British Columbia
Lead organiser. Works on agentic AI, reinforcement learning, and benchmarks for research agents, including large-scale agent evaluation and execution environments for long-horizon research tasks. Responsible for overall coordination, OpenReview setup, the review process and this website.
Co-organizers to be announced
Four to five further organiser slots are open and will be named in the submitted proposal. ICLR permits a maximum of eight organisers. We are recruiting against explicit profiles rather than a rolodex, with balance requirements attached: at least three scientific disciplines, at least three countries, a mix of academia, industry and non-profit, a mix of seniority, gender balance across the final list, and no more than two organisers from any one institution.
- A computational biologist with autonomous-lab or lab-in-the-loop experience
- A mathematician or formal-methods researcher active in autoformalisation or LLM-assisted mathematics
- A computational chemistry or materials researcher working on inverse design or agentic simulation
- An evaluation and benchmarks researcher for scientific agents, to own the reproducibility track
- An AI safety, epistemics or science-policy researcher, to own the safety session and the dual-use review policy
We will also recruit a program committee of roughly 40–60 reviewers, with an open sign-up advertised at least six weeks before the deadline and an explicit call for junior reviewers paired with senior mentors.
Field notes · 10
FAQ
Is this workshop confirmed?
No. FAR is a proposal in preparation for the ICLR 2027 Call for Workshops, which had not been published as of August 2026. Nothing on this page should be read as acceptance, and no speaker listed here has been approached or has agreed to participate. ICLR 2026 accepted 40 of 135 reviewed workshop proposals, so the base rate is roughly 30%.
How is this different from robot-lab and self-driving-lab workshops?
Deliberately and substantially. AutoResearch 2026 at NeurIPS is genuinely half-embodied — robot scientists and physical laboratory automation are one of its core pillars, and it runs AI-authored and AI-review tracks. We consider that valuable and do not intend to duplicate it; we will cross-promote it and point robotics-heavy submissions there.
FAR is robot-light and validation-heavy. Self-driving labs appear here as one verification tier among several — reviews place them at autonomy Level 2–3 of 5, i.e. closed-loop optimisation within a fixed task. Our centre of gravity is the cognitive frontier: conjecture, experiment design, analysis and, above all, verification of claims, including in disciplines where no robot is involved at all. We also do not run an AI-authored paper track: our position is that the field's binding problem is validating AI-generated claims, not staging their authorship.
What is your policy on AI-assisted submissions?
Disclosure, not prohibition. AI assistance in writing, coding, analysis and ideation is permitted and expected, and must be disclosed in a required one-paragraph AI involvement statement that does not count toward the page limit, specifying which systems were used and their versions, the nature of the AI contribution mapped to the research stages it touched, and the human oversight applied — including who verified each empirical claim and how.
Work reporting results produced substantially by an autonomous system must additionally state an autonomy level and, where possible, supply execution transcripts. Fabricated citations or unverified generated results are grounds for desk rejection. Per ICLR policy, papers in the short/position format must be primarily human-authored. Reviewers may use LLM assistance to improve the clarity of their reviews but must not delegate the judgement or upload submissions to services that retain data.
Is the workshop archival? Can I submit work published elsewhere?
Non-archival. Accepted papers are posted on OpenReview with author consent and may be submitted elsewhere afterwards. Work already published at ICLR or another major machine-learning conference is out of scope, following ICLR's non-archival norm; concurrent submission elsewhere is permitted and must be disclosed at submission.
Do I need to be a machine-learning researcher to submit?
No — and we would rather you were not, exclusively. We explicitly solicit work from mathematicians, physicists, chemists, materials scientists and biologists, and reviewers are instructed not to penalise submissions for unfamiliar disciplinary conventions or for lacking ML-benchmark framing. Submissions are assigned so that at least one domain-matched reviewer reads each paper. Rigorous negative results are welcome and have their own award.
Will there be virtual attendance or recordings?
ICLR workshops are in-person only, so the day itself is in the room. Everything we can make durable, we will: accepted papers, posters and slides posted here and on OpenReview and kept online; talks recorded where speakers consent and posted within four weeks; a short public write-up of the panel and safety session within a month. Authors or speakers prevented from travelling — for example by visa refusal — will be accommodated case by case via a pre-recorded talk with live remote Q&A where feasible, or a co-author proxy presenter.