Treat an agent's plan as provisional and replace only the suffix runtime evidence invalidates, instead of replanning from scratch

arXiv 2609.05019·medium signalXBlueskyLinkedInCopy link

TROVE distills evaluated workflow-search traces offline into atomic and composite skills plus an outcome-conditioned transition graph, then at runtime commits one top-level skill at a time and either keeps the remaining route, inserts a trace-supported local response, or replaces only the invalid suffix. Across code generation, QA and math benchmarks with several LLM backbones it beats dataset-level optimization, query-level architecture selection and graph-constrained scheduling on the quality-efficiency trade-off. Ablations are the useful part: composite skills carry most of the offline benefit, insertion supplies the local correction, and suffix replacement is mainly an efficiency win, with early termination doing the heavy lifting on near-saturated tasks.