Prospective Hindsight: Self-Calibrating Reinforcement Learning via Prediction-Reality Gaps

• 类型：arxiv
• 标识：2610.02740
• 链接：https://arxiv.org/abs/2610.02740
• 主分类：agent
• 形态：method
• TLDR：Reinforcement learning for long-horizon agents relies on purely retrospective training signals: credit is assigned only after observing environmental consequences, leaving the agent's belief at action time invisible to the gradient. We introduce Prospective Hindsight (PH), a self-calibrating training principle that augments any retrospective base method with a signal derived from the gap between the agent's prospective prediction (before feedback) and the retrospective evaluation (after feedback). This per-rollout surprise identifies samples where the agent's self-model is most inaccurate and
• 副分类：engineering
• 待LLM分类：否
• 来源文件：
• /inbox/tom/_candidates/2026-10-14-agent-rag-longcontext-candidates.json