A coding agent can explain what went wrong and still repeat the same mistake. The RIPPLE paper calls this the Critique-to-Action Handoff Gap and studies training a single policy that can recover during longer repository-level tasks.
Teach from the evidence available so far
RIPPLE uses a prefix-only teacher to provide recovery supervision. The teacher has no access to future actions or terminal outcomes. After a distillation stage, reinforcement learning assigns separate credit to triggering reflection, diagnosing the failure and taking corrective action.
This design tries to distinguish a useful recovery decision from an explanation that merely sounds plausible after the result is known. For an evaluation, the relevant question is what evidence existed at the moment the agent decided to change course.
Reward the stage that actually helped
Local validity signals gate trajectory advantages so intermediate recovery decisions do not automatically inherit the final outcome. The paper uses controlled ablations to examine stage-aligned credit and matched-prefix interventions to test whether the content of a critique changes the next correction.
Across four repository-level software-engineering benchmarks, RIPPLE reports an arithmetic-mean improvement of 4.8 score points over the base policy. It requires no auxiliary critic at deployment. That is a reported average across those benchmark scores, rather than a universal percentage improvement or an independently measured saving in developer time.
Recovery has a practical boundary
The authors say gains concentrate where the observed prefix contains enough diagnostic evidence and the repository remains reversible within the remaining budget. A model cannot recover from information it never received simply because it has learned to produce a critique.
A reproduction should therefore include cases with missing diagnostics, insufficient time and changes that are difficult to undo. Preserve the intermediate state and check the corrective action against the task’s actual verifier. The contribution is a training approach to recovery; success still depends on the evidence, state and budget of the task being attempted.
Comments