通过分块SVD直接从权重中提取可解释的机制单元，揭示模型内部的触发-写入规则。
Finding Usable Weight Mechanisms with Tiled SVD
- 用分块奇异值分解从线性层权重中提取触发、写入和强度三元组机制。
- 在Gemma-2-2B上，所有7个线性映射均达到182/182的评估通过率，残差写入路径表现最优。
- 适合关注模型可解释性、权重机制挖掘的研究者，代码与工具已开源。
当前主流的机制可解释性方法依赖于训练代理字典（如稀疏自编码器），并基于最大激活文本标注特征。然而，这些概念身份存在于学习到的字典中，而非网络权重本身。本文提出通过列分块奇异值分解（column-tiled SVD）直接从线性层权重中提取机制单元：每个单元为三元组 (v, u, σ)，分别表示触发信号、写入方向和强度。身份由权重规则决定。我们使用预注册评估套件，以完整写入能量提升为标准，而非局部块内提升。在Gemma-2-2B与WikiText-2（16,384词子样本）上，全部7个线性映射被评分：残差写入路径（mlp.down, attn.o）在后子层RMSNorm之后获得完整A/B/C评分，通过52/52个站点层；其余映射（mlp.gate/attn.q/attn.k/effective mlp.up/attn.v）各得A/B评分，26/26通过。总评182/182达标。我们开源了库代码、语料构建工具、实验入口及单元测试。
原文摘要 · Abstract (English)
The dominant approach to mechanistic interpretability trains proxy dictionaries such as sparse autoencoders and labels features from max-activating text. The best such atlases identify con- cepts, but that identity lives in the learned dictionary rather than in the network weights them- selves. We propose extracting mechanism mounts directly from linear sites by column-tiled SVD: each mount is a triple (v,u,σ) read as trigger, write, and strength. Identity is the weight rule. We evaluate mounts with a pre-registered suite judged on full-write energy lift rather than tile-local lift. On Gemma-2-2B with WikiText-2 (16,384-token subsample), all seven linear maps are scored: residual writes (mlp.down, attn.o) receive full A/B/C with steer after post-sublayer RMSNorm and pass 52/52 site-layers; other maps receive A/B only (mlp.gate/attn.q/attn.k/effective mlp.up/attn.v 26/26 each). Aggregate: 182/182 GO. We release library code, the corpus builder, the experiment entrypoint, and unit tests.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。