Research Papers 论文研究 1d ago Updated 20h ago 更新于 20小时前 48

Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability 机制断层扫描:面向控制的可解释性设计测量

Mechanistic tomography unifies patching, gradients, Hessian-vector products, and subset interventions under a single measurement framework (y = Ax + w) for recovering internal model mechanisms A practical iterative procedure is proposed: start with least costly measurements, validate on held-out interventions, calibrate simple mismatches, and expand the measurement family only when structured residuals persist Control provides a demanding validation setting where observer error directly impacts 提出"机制断层扫描"统一框架,将patching、梯度、Hessian向量积、子集干预等测量方法形式化为y=Ax+w的线性测量结构 建立迭代校准流程:从低成本测量起步,在保留干预上测试,校准简单不匹配,当存在结构化残差时扩展测量族 控制场景作为严格验证标准,揭示估计质量与干预效果的关联:控制误差随观测误差上升,目标改进可能掩盖干扰状态移动 不同访问权限下测量策略各异:前向访问用稀疏聚合测量比坐标patching更高效,梯度访问可改进局部归因映射 实证验证:GPT-2-small IOI中Name Mover-Negative Name Mover交互是最大跨组预测项;Qwen-2.5-7B有限

62
Hot 热度
76
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Mechanistic tomography unifies patching, gradients, Hessian-vector products, and subset interventions under a single measurement framework (y = Ax + w) for recovering internal model mechanisms
  • A practical iterative procedure is proposed: start with least costly measurements, validate on held-out interventions, calibrate simple mismatches, and expand the measurement family only when structured residuals persist
  • Control provides a demanding validation setting where observer error directly impacts intervention effectiveness, revealing that target improvement can mask nuisance-state movement
  • Empirical results on GPT-2-small IOI identify the Name Mover-Negative Name Mover interaction as the largest held-out predictive term among cross-group pairs
  • On Qwen-2.5-7B, finite calibration makes additive refusal-response maps adequate, suggesting pairwise lifting may be unnecessary in some practical settings

Why It Matters

This work provides a unified theoretical framework that bridges disparate mechanistic interpretability techniques, enabling researchers to systematically choose and combine measurement methods based on access assumptions and computational constraints. For AI practitioners building controllable systems, the control-oriented validation approach offers a rigorous way to test whether interpretability estimates actually translate to effective intervention, addressing a critical gap between explanation and actionability.

Technical Details

  • Measurement formulation: All interpretability methods are cast as linear measurements y = Ax + w, where A encodes the intervention design, x is the target mechanistic map, and w captures nonlinearities, sampling error, and basis misspecification
  • Iterative calibration procedure: Begin with inexpensive measurements (e.g., sparse aggregate under forward-only access), test predictions on held-out interventions at target scale, calibrate simple mismatches, and only expand to more expensive measurements (lifted measurements, Hessian-vector products) when structured residuals remain
  • Control-theoretic validation: Demonstrates through a two-HMM model that control error rises with observer error, and improvements in target metrics can conceal unwanted movement in nuisance states
  • Empirical benchmarks: On GPT-2-small IOI task, the Name Mover × Negative Name Mover interaction term emerges as the dominant held-out predictive component; on Qwen-2.5-7B refusal-response modeling, finite calibration suffices for additive maps without requiring pairwise lifting
  • Tracr analysis: Shows that the required measurement family is basis-dependent, informing practical choices about representation selection

Industry Insight

  • Organizations investing in mechanistic interpretability should adopt iterative, cost-aware measurement strategies rather than committing to expensive techniques upfront; the framework provides clear decision criteria for when to escalate measurement complexity
  • Control-oriented validation should become a standard benchmark for interpretability methods, as it exposes the critical gap between explanatory accuracy and intervention effectiveness that purely correlational evaluations miss
  • The finding that additive models can suffice for some tasks (Qwen-2.5-7B refusal) while interactions dominate others (GPT-2 IOI) suggests practitioners should empirically determine model-specific complexity requirements rather than assuming uniform interaction structures across architectures

TL;DR

  • 提出"机制断层扫描"统一框架,将patching、梯度、Hessian向量积、子集干预等测量方法形式化为y=Ax+w的线性测量结构
  • 建立迭代校准流程:从低成本测量起步,在保留干预上测试,校准简单不匹配,当存在结构化残差时扩展测量族
  • 控制场景作为严格验证标准,揭示估计质量与干预效果的关联:控制误差随观测误差上升,目标改进可能掩盖干扰状态移动
  • 不同访问权限下测量策略各异:前向访问用稀疏聚合测量比坐标patching更高效,梯度访问可改进局部归因映射
  • 实证验证:GPT-2-small IOI中Name Mover-Negative Name Mover交互是最大跨组预测项;Qwen-2.5-7B有限校准使加法拒绝响应映射已足够,无需成对提升

为什么值得看

本文为机制可解释性研究提供了统一的测量理论框架,将分散的测量技术(patching、梯度、HVP等)纳入同一数学语言,有助于研究者系统选择和应用测量方法。控制验证视角的引入为可解释性估计提供了更严格的评估标准,对开发可靠干预策略具有重要参考价值。

技术解析

  • 统一测量框架:将机制可解释性中的各类测量(patching、梯度、Hessian向量积、子集干预)统一表述为y=Ax+w形式,其中A描述干预设计,x为目标映射,w包含非线性响应、采样误差和基函数误设
  • 迭代校准流程:提出实用程序——从最廉价测量开始,在目标规模的保留干预上测试,校准简单不匹配,当存在结构化残差时扩展测量族
  • 控制验证理论:在双HMM模型中证明控制误差随观测误差单调上升,且目标改进可能隐藏干扰状态移动,揭示"估计引导干预"作为观测者的特殊挑战
  • 不同访问权限下的测量策略:前向访问下稀疏聚合测量可用少于坐标patching的干预数恢复有限效应映射;梯度访问下有限探针改进局部归因映射;提升测量和HVP可恢复一阶映射遗漏的交互
  • 实证结果:GPT-2-small IOI任务中,Name Mover-Negative Name Mover交互是三个测试跨组对中最大的保留预测项;Qwen-2.5-7B上有限校准使加法拒绝响应映射已充分,保留误差不支持成对提升

行业启示

  • 机制可解释性研究应从"单一测量技术"转向"测量族设计"思维,根据模型访问权限(前向/梯度/完整)选择最优测量组合,而非盲目使用坐标patching
  • 控制验证应成为可解释性估计的黄金标准,研究者需警惕"目标指标改进"可能掩盖内部状态异常移动的风险
  • 对于大规模模型(如7B级),加法近似在有限校准后可能已足够,无需过度追求高阶交互恢复,应优先保证测量效率与实用性的平衡

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 LLM 大模型 Alignment 对齐 Evaluation 评测