做 LLM 蒸馏的团队终于有了理论指导——Bridge-Garden 理论解释了为什么混合标签有效，并且直接给出了可落地的方案，训练成本还降了 9.7 倍，建议做模型压缩的开发者点开看看。
这篇论文发现，在 LLM 知识蒸馏中，混合使用教师模型的硬标签（采样 token）和软标签（完整分布）比单独使用任何一种效果更好。作者提出 Bridge-Garden 分解理论，将生成步骤分为“桥”（需精确 token）和“花园”（可灵活选择）两类，硬标签擅长处理桥，软标签擅长处理花园，混合策略能减少训练与推理之间的暴露偏差。基于该理论开发的混合监督方法在 7 组师生模型（含 Qwen、Llama、Gemma、DeepSeek）上优于现有基线，同时将训练成本降低 9.7 倍。代码已开源。
The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works
Knowledge distillation (KD) transfers knowledge from a large teacher model to a smaller student. In language modeling, the student is trained either on tokens sampled from the teacher (hard labels) or the teacher's full next-token distribution (soft labels). Despite soft labels appear strictly richer, we find that mixing hard and soft labels consistently yields better results. Crucially, we show that this gain cannot be explained by closer teacher matching during training. Instead, it comes from reduced exposure bias, the mismatch between training and inference distributions. To explain this phenomenon, we introduce the Bridge-Garden Decomposition theory, which categorizes generation steps into two types: Bridges, where the next token must be exact, and Gardens, where it can be flexible. We show that hard-only KD excels in Bridges by avoiding risky deviations, while soft-only KD preserves diversity in Gardens. A hybrid strategy handles both cases and, as a result, reduces exposure bias across the sequence. Guided by this theory, we develop a family of Bridge-Garden hybrid supervision methods that adaptively balance hard and soft labels. Across a primary suite of seven teacher-student pairs (including Qwen, Llama, Gemma, and DeepSeek) and benchmarks in reasoning and coding, our approach outperforms divergence-based and on-policy KD baselines while reducing training cost by 9.7x, enabling efficient model compression. Code is available at https://github.com/ghwang-s/bridge_garden_hybrid_kd_release.