Multi-Turn On-Policy Distillation with Prefix Replay

• 类型：arxiv
• 标识：2607.04763
• 链接：https://arxiv.org/abs/2607.04763
• 主分类：agent
• 形态：method
• 被引：6
• 被引来源：Semantic Scholar
• S2被引：6
• 影响力被引：0
• TLDR：ReOPD turns expensive agent-environment interaction into a reusable offline resource, enabling scalable distillation across tools, tasks, and environments and preserves or improves OPD-level accuracy, uses zero tool calls during student training, and is at least 4$\times faster per rollout than OPD.
• 待LLM分类：否
• 标题中文：多轮在策略蒸馏与前缀回放
• TLDR中文：ReOPD 将昂贵的 Agent-环境交互转化为可复用的离线资源，实现跨工具、任务和环境的可扩展蒸馏，在学生训练期间保持或提升 OPD 级别的准确率，零次工具调用，并且每次 rollout 至少比 OPD 快 4×。
• 来源文件：
• /inbox/tom/_candidates/2026-07-24-agent-rag-longcontext-candidates.json
• /inbox/tom/_candidates/2026-07-25-agent-rag-longcontext-candidates.json
• /inbox/tom/_candidates/2026-07-26-agent-rag-longcontext-candidates.json
• /inbox/tom/_candidates/2026-07-27-agent-rag-longcontext-candidates.json
• /inbox/tom/_candidates/2026-07-27-agent-memory-tool-use-candidates.json
• [S2 enrich]