跳转至

文章背景与核心概要

传统的机械可解释性（Mechanistic Interpretability）研究通常依赖于训练代理字典（例如稀疏自编码器 SAE），并通过最大激活文本来对特征进行标注。尽管这些方法在概念识别方面十分有效，但这些表征通常存在于所学习到的字典中，而非模型的核心权重之内。

本文引入了一种直接的方法：利用列分块奇异值分解（column-tiled SVD）直接从线性位点中提取机制挂载点（mechanism mounts）。每个提取出的挂载点都表示为一个三元组 \((v, u, \sigma)\)——分别对应触发器（trigger）、写入器（write）和强度（strength），其身份直接由权重规则定义。通过在 Gemma-2-2B 模型上使用 WikiText-2 数据集进行全写入能量提升（full-write energy lift）评估的预注册套件测试，该方法成功对所有线性映射进行了评分，在所有测试层和配置中获得了累计 182/182 的“通过（GO）”评级。

Finding Usable Weight Mechanisms with Tiled SVD

Authors: Ash Manvi, Samreena Tajreen
Published: August 7, 2026
arXiv: 2608.06969 [cs.AI]

Authors: Ash Manvi, Samreena Tajreen
Published: August 7, 2026
arXiv: 2608.06969 [cs.AI]

Summary

Traditional mechanistic interpretability often relies on training proxy dictionaries (such as sparse autoencoders) and labeling features via max-activating text. While effective for concept identification, these representations exist within the learned dictionary rather than the model's core weights.

This paper introduces a direct approach: extracting mechanism mounts straight from linear sites using column-tiled SVD. Each extracted mount is represented as a triple \((v, u, \sigma)\)—corresponding to a trigger, write, and strength—where identity is defined directly by the weight rule. Evaluated via a pre-registered suite measuring full-write energy lift on Gemma-2-2B using WikiText-2, the method successfully scores all linear maps with a cumulative 182/182 "GO" rating across all tested layers and configurations.

Summary

Traditional mechanistic interpretability often relies on training proxy dictionaries (such as sparse autoencoders) and labeling features via max-activating text. While effective for concept identification, these representations exist within the learned dictionary rather than the model's core weights.

This paper introduces a direct approach: extracting mechanism mounts straight from linear sites using column-tiled SVD. Each extracted mount is represented as a triple \((v, u, \sigma)\)—corresponding to a trigger, write, and strength—where identity is defined directly by the weight rule. Evaluated via a pre-registered suite measuring full-write energy lift on Gemma-2-2B using WikiText-2, the method successfully scores all linear maps with a cumulative 182/182 "GO" rating across all tested layers and configurations.

Key Details

  • Primary Subject: Artificial Intelligence (cs.AI)
  • Methodology: Column-tiled Singular Value Decomposition (SVD) for direct weight mechanism extraction.
  • Evaluation Framework: Pre-registered suite judged on full-write energy lift rather than tile-local lift.
  • Target Model & Corpus: Gemma-2-2B evaluated on a 16,384-token subsample of WikiText-2.
  • Results Overview:
  • Residual Writes (mlp.down, attn.o): Received full A/B/C ratings (with steering after post-sublayer RMSNorm) and passed across all 52/52 site-layers.
  • Other Maps (mlp.gate, attn.q, attn.k, mlp.up, attn.v): Received A/B ratings (26/26 each).
  • Aggregate Score: 182/182 GO.

Key Details

  • Primary Subject: Artificial Intelligence (cs.AI)
  • Methodology: Column-tiled Singular Value Decomposition (SVD) for direct weight mechanism extraction.
  • Evaluation Framework: Pre-registered suite judged on full-write energy lift rather than tile-local lift.
  • Target Model & Corpus: Gemma-2-2B evaluated on a 16,384-token subsample of WikiText-2.
  • Results Overview:
  • Residual Writes (mlp.down, attn.o): Received full A/B/C ratings (with steering after post-sublayer RMSNorm) and passed across all 52/52 site-layers.
  • Other Maps (mlp.gate, attn.q, attn.k, mlp.up, attn.v): Received A/B ratings (26/26 each).
  • Aggregate Score: 182/182 GO.