Station agents scored 62.7% in a research rediscovery test, authors report

Wenyu Du and Stephen Chung report that AI agents in Station, a multi-agent research environment, recovered an average of 62.7% of criteria drawn from the findings of three ICLR oral papers. Their preprint, submitted to arXiv on 6 October 2026, tests whether agents can pursue research questions without an intermediate score to guide them.

The percentage needs a careful reading: it measures coverage of findings broken into individual criteria, not the share of papers reproduced in full. The captured abstract does not identify the papers or explain how the criteria were selected and graded. It also leaves out run counts and computing budgets—details needed to assess how much weight to put on the average.

Station is described as an open-world environment where multiple agents simulate a scientific ecosystem. For this study, the authors added a Supervisor mechanism and periodic Meta Reflection, both intended to sustain exploration when intermediate metrics are unavailable. They report that analyses comparing system variants and examining agent behavior found better research coverage and continuity when the two mechanisms were used together.

In the three-paper test, agents received each paper’s main research question, while its results were withheld and web access was disabled. The authors report lower criteria-rediscovery scores for two comparison systems: 15.4% for Codex Multiagent-v2 and 14.4%–20.6% for AI Scientist-v2. The abstract does not describe how those systems were configured.

The researchers also tried two open-ended tasks without reference papers. They say some agent discoveries closely matched work researchers reported after the knowledge cutoff. That is a separate claim from recovering known findings; the abstract supplies no examples with which to judge how close those matches were.