NeuralStackAutonomous AI research has largely been organized as a series of isolated experiments, with each agent system generating ideas, running tests and writing reports without a durable way to build on earlier work. AgentRxiv proposes a different model: a shared preprint server where LLM-powered research laboratories can publish findings and retrieve the work of other laboratories.
Developed by Samuel Schmidgall of Johns Hopkins University and Michael Moor of ETH Zurich, the framework is designed to make autonomous research cumulative. Agents can use prior reports to develop new reasoning and prompting techniques, while separate laboratories can work in parallel toward a common research objective.
In the study, agents with access to their own earlier research achieved an 11.4% relative improvement over baseline on MATH-500. The strongest strategy also transferred to other domains, producing an average improvement of 3.3% across additional benchmarks and language models.
The collaborative setup produced a larger gain. Multiple agent laboratories sharing research through AgentRxiv achieved a 13.7% relative improvement over baseline on MATH-500 and progressed more rapidly than isolated laboratories. The paper also reports that, using GPT-4o mini as the base model, accuracy rose from 70.2% to 78.2% with the best discovered reasoning technique.
AgentRxiv builds on earlier autonomous research systems, including AI Scientist, Virtual Lab and Agent Laboratory. Those projects demonstrated that language-model agents can generate research ideas, coordinate specialized roles, conduct experiments and produce scientific reports, but the authors argue that they do not adequately support continuous development across successive research efforts.
The framework uses a centralized, open-source archive to let agents share results as soon as papers are submitted. That design supports asynchronous collaboration and allows laboratories to scale research across available computing resources, rather than forcing every experiment into a single sequential workflow.
Why this matters
AgentRxiv shifts the focus of autonomous research from individual agent capability to the quality of the research ecosystem around agents. If systems can reliably discover, evaluate and reuse useful techniques, they could accelerate progress by turning separate model runs into a growing body of machine-generated knowledge. The trade-off is that parallel research can improve wall-clock speed while requiring more computation, and shared outputs will still need careful evaluation before they can be treated as dependable scientific results.
The authors see the framework as a step toward AI systems that help design future AI systems alongside human researchers. Its longer-term significance will depend on whether the gains generalize beyond the tested reasoning tasks and whether future versions can strengthen verification, reproducibility and safeguards around autonomous scientific work.