Tag
This paper introduces RLVP (Reward the Outcome, Penalize the Path), a reinforcement learning method that uses a verifiable penalty for path violations and outcome reward to achieve near-zero constraint violations with high task success, improving sample efficiency in real-world agentic environments.