AI agents learn better by tracking what they believe, not just actions

Original: Rewarding Beliefs, Not Actions: Consistency-Guided Credit Assignment for Long-Horizon Agents

Source: arxiv.org ↗