Decision Shifts, Lost Label Functionality, and an Inconclusive Grounding Audit in Correctness-Gated Multi-Teacher Distillation
作者: Xiaofei Feng
分类: cs.AI
发布日期: 2026-09-09
💡 一句话要点
提出正确性门控多教师蒸馏以解决决策偏移问题
🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture)
关键词: 多教师蒸馏 决策正确性 推理基础 模型优化 实验分析
📋 核心要点
- 现有的多教师蒸馏方法在决策正确性与推理基础之间存在偏移，导致标签功能丧失。
- 本文提出了一种正确性门控的多教师蒸馏方法，通过优化决策过程来提升模型性能。
- 实验结果显示，正确性加权臂在准确率和宏F1上有显著提升，但未能证明系统级的整体改善。
📝 摘要（中文）
本文探讨了候选决策正确性与推理基础的不同目标，分析了固定实验中的正确性门控多教师蒸馏。实验涉及八个臂，共享4,330个源，63.9M参数的学生模型，以及267个保留示例的评估。相较于未过滤的蒸馏，正确性加权臂在准确率、五标签宏F1和条件不安全行动率上表现出显著差异。然而，这些变化并不意味着行为的普遍改善，且在可用性修订审计中，结果显示加权和未过滤输出的支持证据存在显著差异，审计结果无法确认系统级的改进或损害。
🔬 方法详解
问题定义：本文旨在解决多教师蒸馏中决策正确性与推理基础之间的矛盾，现有方法在标签功能丧失方面存在不足。
核心思路：提出的正确性门控多教师蒸馏方法通过引入正确性加权机制，旨在优化模型的决策过程，提升其在特定任务上的表现。
技术框架：整体架构包括多个教师模型与一个学生模型，采用固定实验设计，利用多轮优化和更新来调整学生模型的参数。主要模块包括数据输入、模型训练和结果评估。
关键创新：本文的创新点在于引入了正确性加权机制，强调决策的准确性和推理基础的结合，与传统的蒸馏方法相比，能够更好地处理标签功能丧失的问题。
关键设计：在实验中，设置了多个参数，包括优化行数、更新次数和教师模型的响应池，采用了特定的损失函数来平衡不同目标的权重。
🖼️ 关键图片
📊 实验亮点
实验结果显示，相较于未过滤的蒸馏，正确性加权臂在准确率上提高了0.1660，五标签宏F1提高了0.1323，条件不安全行动率降低了0.4979。然而，审计结果表明加权臂在某些情况下未能提供额外的决策收益。
🎯 应用场景
该研究的潜在应用领域包括自然语言处理、图像识别和其他需要多模型协作的智能系统。通过提升模型的决策正确性和推理能力，能够在实际应用中提高系统的可靠性和安全性，具有重要的实际价值和未来影响。
📄 摘要（原文）
Candidate decision correctness and rationale grounding are different objectives. We examine correctness-gated multi-teacher distillation in a fixed experiment. Eight arms share 4,330 sources, a 63.9M-parameter student, 12,990 optimization rows, 406 updates, evidence inputs, and a decoder; seven teacher-based arms use one fixed three-response pool. Three seeds are evaluated on 267 held-out examples. Relative to unfiltered distillation, the correctness-weighted arm differed in accuracy by +0.1660 (95% observed-matrix interval [0.0670, 0.2455]), five-label macro-F1 by +0.1323 ([0.0916, 0.1731]), and task-defined conditional unsafe-action rate by -0.4979 ([-0.5926, -0.3686]). These shifts do not imply uniformly better behavior. Source-label SFT had the highest mean macro-F1 (0.586). The weighted arm had zero Refuted recall in every seed, and two seeds assigned NotEnoughInfo to all 167 claim examples. In an availability-amended audit at one reference seed, weighted and unfiltered outputs had 0/20 versus 1/20 evidence-supported positives and 20/20 versus 19/20 positives containing unsupported material. Samples were non-paired, source overlap was not serialized, and the amendment followed automatic summarization but preceded annotation. The audit therefore cannot estimate a common-source grounding effect and is inconclusive about system-level improvement or harm. Hard filtering already achieved 0.660 accuracy, 0.530 macro-F1, and 0.135 conditional unsafe rate. The implemented weighted arm showed no demonstrated incremental decision benefit over hard filtering. This fixed-matrix failure analysis shows decision redistribution with lost label functionality; the available human audit does not establish a grounding gain.