AI Agents Replicate All ICML Oral Papers Exposing Low Reproducibility
Reactions from ranked influencers
4 posts
How much of science is verifiable? As AI agents become capable of increasingly complex, long-horizon tasks, we have an opportunity to rethink the research ecosystem. One possibility is to verify the science rather than only review the narrative in a paper, a clear blind spot in human peer review. When ICML wrapped up, we used AI agents to review and replicate all 168 oral papers, including 105 full replications. We ask three questions: • What does it cost to produce a top machine learning paper? • What can replication reveal that narrative-only reviewing misses? • How well does our review system align with existing human reviews? What we found: • The median estimated cost to reproduce every reported experiment was $8,900. • Despite the cost, papers are rarely replicable. Only 7 reproduced more than 80%, and only 27 reproduced more than 40% of the claims we tested. • Our system covered ~80% of the issues identified by two or more human reviewers. Check out the full article. More links in the thread if you are interested! #MachineLearning #AI #Science #PeerReview #Reproducibility #ICML
Full results: https://sai.science/icml You can also try our review service: https://sai.science/
Wish I know gifs this well!
@ChenhaoTan > We replicated ICML 2026 oral papers, and only 7 mostly held up. Replication crisis! Or the agents aren't that good? > We used AI agents to review and replicate all 168 oral papers, including 105 full replications
This project with Chenhao and SAI Labs has been revealing. We replicated all 168 oral papers at the top AI Conference ICML and only 7 completely held up. To be clear, the significance of this demonstration is not that peer review is getting harder with AI (although it is -- Holden Thorp: https://www.science.org/doi/10.1126/science.aek5570), it's that it's NEVER been a sufficiently strong verifier to facilitate cumulative assembly. Verifying from top-to-bottom reveals what's fundamentally challenging about distributed science, and what can be done to enable it, especially when collective advance is the goal. Check out and get a full (including code) free verification / review @ https://sai.science/.
How much of science is verifiable? As AI agents become capable of increasingly complex, long-horizon tasks, we have an opportunity to rethink the research ecosystem. One possibility is to verify the science rather than only review the narrative in a paper, a clear blind spot in human peer review. When ICML wrapped up, we used AI agents to review and replicate all 168 oral papers, including 105 full replications. We ask three questions: • What does it cost to produce a top machine learning paper? • What can replication reveal that narrative-only reviewing misses? • How well does our review system align with existing human reviews? What we found: • The median estimated cost to reproduce every reported experiment was $8,900. • Despite the cost, papers are rarely replicable. Only 7 reproduced more than 80%, and only 27 reproduced more than 40% of the claims we tested. • Our system covered ~80% of the issues identified by two or more human reviewers. Check out the full article. More links in the thread if you are interested! #MachineLearning #AI #Science #PeerReview #Reproducibility #ICML