The research introduces RIPPLE, a system designed to adapt policies in workflow synthesis agents. The system addresses challenges associated with persistent prompt editing, specifically concerning edit locality and composition sensitivity. RIPPLE separates the decision of where to make an edit from the assessment of whether that edit remains safe after the workflow is composed. It diagnoses failed trajectories, maps these failures to specific policy segments, and restricts corrections to those segments.
To evaluate the system, RIPPLE was tested on Flow-HO, a synthetic benchmark for executable workflow synthesis. Results demonstrated an improvement in validation success rates by up to 23.1%. The system also yielded positive gains on two additional frozen language-model backbones. The evaluation focused on maintaining edit efficiency and minimizing execution cost.
Targeted interaction analysis highlighted the system’s key properties. A segment-local tool-use edit demonstrated a change in downstream resource resolution and validation. Conversely, an edit that was beneficial in isolation became detrimental after composition. This illustrates the composition-sensitive nature of the adaptation process.
The RIPPLE approach offers a method for improving agent performance in workflow synthesis by mitigating the risks associated with broad policy changes. The system’s replay-informed strategy allows for a more robust and reliable adaptation process.
Source: https://arxiv.org/abs/2609.12127