HKU's Station agents rediscover 62.7% of ICLR paper findings autonomously

TheHKU · hf · 2026-10-09

HKU researchers introduce Station, an open-world environment where multiple agents simulate a scientific ecosystem, to test whether AI can autonomously pursue open-ended scientific discovery.

To address agents giving up when intermediate metrics are missing, they add a Supervisor mechanism and periodic Meta Reflection to encourage persistent exploration.
Tasks are built from three ICLR oral papers: agents get only the research question (no results, no web access), and the metric is how many original findings they rediscover.
Station rediscovers 62.7% of criteria on average, vs 15.4% for Codex Multiagent-v2 and 14.4–20.6% for AI Scientist-v2; ablations show the two mechanisms jointly improve research coverage and continuity.
On two open-ended tasks without oracle papers, some agent discoveries closely match findings published by researchers after the models' knowledge cutoff.

Original post →

HKU's Station agents rediscover 62.7% of ICLR paper findings autonomously

More from coding & agent