NeuralStackAutonomous AI research systems have largely worked as isolated laboratories, producing papers and experiments without a durable way to build on one another’s findings. AgentRxiv, a framework introduced by Samuel Schmidgall of Johns Hopkins University and Michael Moor of ETH Zurich, proposes a shared preprint server where LLM-based agent laboratories can publish, retrieve, and extend research reports.
The system is designed to make autonomous research cumulative rather than one-off. Agents can use earlier reports to develop new reasoning and prompting techniques, while separate laboratories can pursue the same research goal in parallel and exchange their findings through the platform.
In experiments focused on reasoning strategies, agents with access to their own prior research achieved an 11.4% relative improvement over an isolated baseline on MATH-500. The best-performing strategy also transferred to other domains, producing an average improvement of 3.3% across additional benchmarks and models, according to the researchers.
Collaboration between multiple agent laboratories produced a larger gain. The researchers report a 13.7% relative improvement over baseline on MATH-500 when laboratories shared their work through AgentRxiv, while parallelized research accelerated progress in elapsed time.
The results build on earlier autonomous research efforts, including AI Scientist, Virtual Lab, and Agent Laboratory. Those systems demonstrated that LLM agents could generate ideas, write research code, run experiments, incorporate feedback, and produce scientific reports, but the authors argue that they lacked a mechanism for sustained, cross-laboratory accumulation of knowledge.
AgentRxiv treats research outputs as reusable artifacts rather than final products. Its centralized, open-source design is intended to let each generation of agents inspect previous discoveries, test new approaches, and contribute improved results back to the shared archive.
Why this matters
If the approach holds beyond the reported evaluations, it could shift autonomous AI research from a collection of disconnected experiments toward a network of cooperating laboratories. That would give organizations a way to scale research across multiple agent systems while preserving a record of ideas and results that later agents can reuse.
The trade-off is that faster parallel progress can require more computational resources, and a shared archive does not by itself guarantee that an agent’s findings are correct or scientifically meaningful. Human researchers would still need to provide research direction, evaluate evidence, and determine which discoveries merit further investigation.
AgentRxiv’s broader proposal is to make collaboration a core capability of research agents. By enabling systems to share reports and iteratively improve one another’s methods, the framework suggests a path toward AI tools that participate in designing future systems alongside human scientists.