A new research agent called PrimeScientist decides not just what experiment to try next, but how much of its shrinking budget that experiment deserves.
Researchers built PrimeScientist to treat resource allocation as part of the research problem itself, not an afterthought. The system keeps a tree of competing research plans and their outcomes, then applies a Monte Carlo tree search policy that weighs unexplored directions against ones already showing promise while tracking remaining budget. It was tested across AI research, systems and code optimization, and machine-learning engineering tasks, letting the agent reallocate effort mid-run instead of locking into a fixed plan upfront. Across 12 AI research tasks, PrimeScientist beat a baseline called AutoResearch by 10.3% on average reward while using 50.6% fewer research attempts under the same budget.
That efficiency gain matters more than the reward number. Autonomous research agents are already good at generating ideas; the bottleneck is compute and time to test them, and most systems still burn through both on weak plans before abandoning them. Making that spending decision explicit and optimizable, rather than fixed in advance, is a modest design change that could stretch a fixed compute budget considerably further.
Every automated-science pitch promises efficiency; this one at least backs the claim with a named baseline and a fixed budget, rather than just a headline number.