AutoScientists: decentralized agents top BioML-Bench baselines

AutoScientists, a decentralized team of AI agents described in a paper posted to arxiv on May 27 by Shanghua Gao, Ada Fang, and Marinka Zitnik, reports a 74.4% mean leaderboard percentile across 24 tasks on BioML-Bench, an improvement of +8.33 points over the strongest AI agent the authors benchmark against.

The pitch is a shot at the orchestrator pattern. Existing AI-research agents, the paper argues, "typically follow a single research trajectory or coordinate through a central planner with fixed objectives." AutoScientists has no central planner. Instead, "Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before using experimental compute," and share both successes and failures to cut down on redundant exploration.

The numbers lean on three benchmarks. On a GPT training-optimization task, the paper claims AutoScientists "reaches a target validation bits-per-byte 1.9x faster than Autoresearch" and keeps finding improvements where a single-agent approach stalls out, 7 accepted improvements versus 0. On the ProteinGym assay for ACE2-Spike binding, it reports a method that "improves over the current state-of-the-art model by +12.5% in Spearman correlation"; the same method applied unchanged across all 217 ProteinGym assays scores +6.5%. The abstract does not name the LLM powering the agents or the absolute compute budget behind the 'matched' comparisons.

Code is on GitHub. Two AI researchers we follow in our Who's Who directory circulated the paper link.