Reproduce “Stateless Language Agents: Scaling Long-Horizon Automated Research” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qizheng Zhang, Changxiu Ji, Isaac Sun, Yuetai Li, Shubhangi Upasani, Sherry Ruan, Boyuan Ma, Fenglu Hong and 10 more
SLA stores durable research state in a harness, while fresh agents receive focused contexts and Advisor assignments to parallel Workers. The harness retains candidate solutions and measured outcomes across invocations, but not agent conversations. On the Anthropic kernel task, SLA reaches SwarmResearch’s mean final performance after 67.9M cumulative tokens, versus 986.3M; most other task and model configurations use single runs.
Key insights
- Retaining research state does not require retaining agent conversations: controlled context reconstruction can make accumulated evidence usable without repeatedly replaying history.
- Long-horizon results can reverse short-run rankings: SLA trails leading baselines on all four FrontierSWE tasks at one-quarter budget but leads them at full budget.
- Coordination and isolation help prevent duplicated work, while the best Worker count and model strength depend on the task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.