PRIMESCIENTIST 用树状分配实验预算，让科研智能体少做无效尝试
这篇讲的是怎么让科研智能体把实验预算花在刀刃上，同预算下比 AutoResearch 奖励高 10.3%，尝试次数省一半，做法挺巧。
PRIMESCIENTIST 是一个科研智能体框架，核心思路是让智能体自己决定实验预算花在哪些方向，而不是一条路径跑到底。它同时维护多个可执行方案的树结构，实验结果会更新各分支的价值评估，预算充裕时多探索、预算减少时集中到强分支。在 12 个 FIRE-Bench 任务上，平均奖励比 AutoResearch 高 10.3%，研究尝试次数少 50.6%，token 预算保持相同。在全部 24 个任务的评估里，它在其中 23 个上用了更少的尝试次数。
PRIMESCIENTIST shows that research agents should choose where to spend their experiments, rather than keep pushing the same idea until the budget runs out.
Instead of following a single trajectory, it keeps competing executable plans in a tree. Experiment results update branch values, while the allocation policy explores more when resources are plentiful and concentrates on stronger branches as the budget shrinks.
On 12 FIRE-Bench tasks, it delivered 10.3% higher average reward than AutoResearch while using 50.6% fewer research attempts under the same token budget. Across the full evaluation, it used fewer attempts on 23 of 24 tasks.
