Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

2026-10-06

AI Summary

This study addresses the absence of learning signals caused by reward indifference in reinforcement learning, where agents struggle to extract supervision from failed trajectories. To overcome this, we propose a prospective learning framework centered on a novel Self-Retrospection Distillation method. This approach leverages hindsight experience to guide foresight prediction, distilling privileged information into prospective targets for blind policies. By integrating verifiable reward reinforcement learning with self-distillation, it enables knowledge transfer under zero-reward contrast scenarios without requiring prospective signal generation at inference time. Experimental results demonstrate that our method improves success rates on long-horizon tasks by up to 24.2 percentage points. Notably, in settings with 98% complete failure, the success rate increases dramatically from 0% to 60.6%.

Abstract

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and how the agent fails. We ask a complementary question: can hindsight teach an agent what it could have anticipated before acting? We introduce prospective learning, which uses post-hoc experience to supervise foresight predictions from the pre-interaction view, and instantiate it with Self-Retrospection Distillation (SRD). Intuitively, a completed trajectory reveals knowledge that would have been useful and pitfalls that should be avoided; SRD distills this privileged hindsight into trajectory-blind foresight of the same policy. Foresight serves only as a training target and need not be explicitly generated at inference time. Across 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines with gains of up to $24.2$ pp. Its advantage is especially pronounced when reward contrast is scarce: when $37$--$98\%$ of rollout groups are reward-uniform across model scales, yet SRD can still exploit learning signal from sampled trajectories. In the 2B setting, where $98\%$ of groups are all-failure, the RLVR training ends up at $0.0\%$ success, while adding SRD reaches $60.6\%$ under the same rollout budget. Our results suggest that post-hoc agent experience is useful not only for evaluating or improving behavior, but also for shaping predictive representations before available interaction.

Problem

• Reinforcement Learning with Verifiable Rewards
• Reward Sparsity
• Group-relative Objectives
• Hindsight Learning

Innovation

• Self-Retrospection Distillation
• Prospective Learning
• Reinforcement Learning with Verifiable Rewards
• Hindsight-to-Foresight
• Reward Contrast