Autonomous research agents can generate more experimental ideas than they have time or computing budget to test. PrimeScientist addresses that bottleneck by deciding both which direction to pursue and how many resources to invest after each result.
The system represents competing hypotheses and their outcomes in an executable plan tree. A policy based on Monte Carlo Tree Search balances exploring a new branch against refining an approach that has already produced promising evidence. Crucially, the amount of budget remaining is part of every decision rather than an afterthought.
Across 12 AI research tasks, the authors report that PrimeScientist improved average reward by 10.3% compared with AutoResearch while requiring 50.6% fewer attempts. Additional evaluations covered systems work, code optimization and machine-learning engineering, where the same allocation strategy improved research quality and sample efficiency together.
These are benchmarked agent workflows, not evidence that the system can independently conduct open-ended science. The result targets a practical weakness in current automation: proposing an experiment is cheap, while running, interpreting and following up on it can be expensive. Keeping failed and competing plans visible may also reduce repeated work. Real deployments would still need human judgment over scientific validity, safety and whether a numerical reward captures a worthwhile research goal.