What they did
Offline, the authors distilled evaluated workflow-search traces into atomic skills, composite multi-step skills, and a transition graph conditioned on observed outcomes. This representation separates stable workflow fragments from decisions that may need to change during execution.
Online, TROVE executes one top-level skill at a time and then checks the remaining route against the resulting evidence. It can retain the continuation, insert a trace-supported response, or replace only the invalid suffix. The authors evaluated the approach on code-generation, question-answering, and mathematical-reasoning benchmarks using multiple language-model backbones and compared it with dataset-level workflow optimization, per-query architecture selection, and graph-constrained scheduling.
Key findings
- TROVE produced a better reported quality–efficiency trade-off than the tested orchestration baselines across the three task families.
- Quality gains were largest on tasks where intermediate outcomes changed which continuation was appropriate.
- Early termination reduced computation substantially on tasks where performance was already near saturation.
- Ablations attributed most offline gains to composite skills; insertion supported local correction, while suffix replacement mainly improved efficiency.
Why it matters
The results support selective route editing as an alternative to both rigid preplanned workflows and wholesale replanning. For agent systems whose intermediate results reveal whether later steps remain useful, preserving validated work could reduce wasted model calls while retaining the ability to adapt.
Caveats
The provided abstract does not report benchmark names, absolute scores, computational costs, statistical uncertainty, or the scale and expense of offline workflow search. It therefore does not establish how broadly the gains transfer beyond the evaluated task families, nor whether trace-derived skills remain effective under distribution shifts or with substantially different tools and models.