Search papers, labs, and topics across Lattice.
This paper introduces PBSD (Privileged Bayesian Self-Distillation), a novel method designed to tackle the credit assignment problem in long-horizon reinforcement learning tasks by providing fine-grained feedback on intermediate reasoning steps. By leveraging Bayes' rule, PBSD converts the challenging posterior-to-prior probability ratios into actionable turn-level credit signals, allowing agents to discern which actions contribute positively or negatively to the final outcome. Experimental results show that PBSD significantly enhances performance in both in-domain and out-of-domain scenarios, indicating its effectiveness in improving policy learning and generalization across varying contexts.
Sparse rewards can be transformed into actionable turn-level feedback, enabling agents to learn from both successful and misleading actions in long-horizon tasks.
Long-horizon agentic tasks pose a fundamental credit assignment challenge for outcome-base reinforcement learning: trajectory-level rewards verify final correctness but provide limited guidance on which intermediate reasoning steps or tool interactions contribute to the outcome. The difficulty is especially pronounced in multi-turn search agents, where successful trajectories may contain misleading actions and failed trajectories may contain valuable evidence-gathering steps. We propose PBSD (Privileged Bayesian Self-Distillation), a Bayes-calibrated self-distillation method for fine-grained credit assignment under sparse final rewards. PBSD measures trajectory quality through the posterior-to-prior probability ratio of the verified answer and applies Bayes' rule to convert this hard-to-estimate answer-side ratio into a tractable likelihood ratio between a standard student model and a privileged answer-conditioned teacher model. Autoregressive decomposition of this Bayesian evidence score yields turn-level signals that identify whether each intermediate turn supports or undermines the verified outcome. Consequently, PBSD provides a principled and elegant reweighting scheme that transforms sparse outcome supervision into Bayes-calibrated turn-level credit signals, while remaining fully compatible with standard policy optimization. Experiments demonstrate that PBSD consistently enhances performance across both in-domain and out-of-domain settings, and effectively transfers knowledge from short-context training to long-context inference, suggesting that its fine-grained credit assignment mechanism facilitates more effective policy learning and yields improved generalization.
