EVENT DOSSIER
The researchers proposed a new method called TROVE, aimed at optimizing agent skill scheduling through runtime evidence verification and editing corrections. In the offline phase, this method distills workflow search traces into atomic and composite skills and constructs a result condition transition graph; in the online phase, it treats the planned route as a temporary solution, retaining valid continuation after confirming the top-level skills, inserting trace-based local responses, or replacing invalid suffixes. Evaluation across code generation, question-answering, and mathematical reasoning benchmarks showed that TROVE offered a better quality-efficiency trade-off compared to baseline methods such as dataset-level optimization, query-level architecture selection, and graph-based scheduling. Ablation experiments confirmed that composite skills captured the main offline benefits, insertion implemented local corrections, and suffix replacement primarily improved efficiency.
提出 Trace-grounded Route Orchestration via Validation and Editing（TROVE），通过运行时证据验证与编辑修正执行路径。该方法离线将工作流搜索痕迹蒸馏为原子及复合技能，并构建结果条件转换图；在线阶段将计划路线视为临时方案，在确认顶层技能后保留有效续行、插入基于痕迹的局部响应或仅替换无效后缀。跨代码生成、问答及数学推理基准的评估显示，TROVE 相比数据集级优化、查询级架构选择及图约束调度等基线，提供了更优的质量 - 效率权衡；当结果改变续行时质量提升最大，而早期终止在近饱和任务上带来显著效率增益。消融实验证实复合技能捕获主要离线收益，插入实现局部修正，后缀替换主要提升效率。