Can We Trust the Teacher? Decoupled Credit Direction-Magnitude for Self-Distillation

• 类型：arxiv
• 标识：2609.34848
• 链接：https://arxiv.org/abs/2609.34848
• 主分类：multimodal
• 形态：method
• TLDR：RLVR provides reliable trajectory-level credit, while OPSD offers dense supervision for token-level credit. This exposes a fundamental coupling when updating step-level credit direction and magnitude with teacher supervision, preventing steps from receiving reliable credit directions and contribution magnitudes, while making both vulnerable to teacher judgment errors and preference variance, as supported by our theoretical analysis. To separate credit direction from its contribution magnitude, we introduce Decoupled Credit Self-Distillation (DCSD), which theoretically decouples credit directio
• 待LLM分类：否
• 标题中文：我们能信任教师吗？面向自蒸馏的解耦信用方向-幅度
• TLDR中文：RLVR 提供可靠的轨迹级信用，OPSD 提供 token 级的稠密监督。这暴露出在教师监督下更新步骤级信用方向与幅度时存在根本性的耦合，使得步骤难以获得可靠的信用方向与贡献幅度，且两者都易受教师判断错误与偏好方差的影响，我们的理论分析证实了这一点。为分离信用方向与其贡献幅度，我们提出解耦信用自蒸馏（DCSD），从理论上解耦信用方向与幅度。
• 来源文件：
• /inbox/tom/_candidates/2026-09-30-agent-rag-longcontext-candidates.json