Pairing outcome rewards with verifiable per-action path penalties reduces constraint violations nearly sixfold at equal task success, while a progress potential accelerates learning only where partial progress is reachable.
The learned critic is a poor training reward in every quadrant; its only robust use is offline diagnosis at the intent frontier
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
RLVP: Penalize the Path, Reward the Outcome
Pairing outcome rewards with verifiable per-action path penalties reduces constraint violations nearly sixfold at equal task success, while a progress potential accelerates learning only where partial progress is reachable.