RIPPLE: an edit confined to one prompt-policy segment changes downstream behavior, so replay candidate edits after previously accepted ones before persisting
Prompt-policy editing improves workflow-synthesizing agents without touching model weights, but the paper names two coupled properties practitioners get wrong: edit locality does not imply effect locality (a segment-local edit ripples into downstream execution), and edit effects are composition-sensitive (edits that help in isolation interfere after composition and can become harmful). RIPPLE separates the two decisions — diagnose a failed trajectory and map the failure to one predefined policy segment, then evaluate candidates against the same iteration-start policy for isolated gain before replaying promising edits on top of already-accepted ones to expose interaction. Only edits that survive composition are retained; validation success improved by up to 23.1% on the Flow-HO benchmark with gains holding on two other frozen backbones.
↳ Follow the thread