arXiv cs.AI· Daking Rai, Mor Geva, Ziyu Yao·· 3 天前AI 评分23
面向语言模型可解释性的数据驱动电路发现（DCD）
Data-driven Circuit Discovery for Interpretability of Language Models
AI 导读
研究提出数据驱动电路发现（DCD），在电路发现流程前加入分组步骤，用逐样本归因模式聚类（DCD-Cluster）将数据集划分为依赖相似机制的多个子集。作者先证明当数据集混合不同机制处理的任务时，现有电路发现会把多种机制合并为单一电路且保持与单独任务相当的忠实度，并在四个任务、三个模型上给出任务内机制多样性的证据。
正文 · 原文
Abstract:Circuit study aims to explain how language models (LMs) implement a task by localizing and interpreting a circuit: a computational subgraph responsible for a specific model's behavior. Existing methods for circuit localization are largely hypothesis-driven: they define a task through a dataset, assume that all examples in the task are implemented by the same mechanism, and apply circuit discovery to obtain a single circuit. This workflow does not account for mechanistic diversity within a task, where different examples may rely on different mechanisms. We first show that when a dataset mixes tasks processed through distinct mechanisms, current circuit discovery can combine these mechanisms into a single circuit while retaining faithfulness comparable to circuits discovered on each task separately. Across four tasks and three models, we further provide evidence for such mechanistic diversity within individual tasks. To account for this diversity, we propose Data-driven Circuit Discovery (DCD), which modifies the circuit-discovery workflow by adding a grouping step before the discovery. The grouping step aims to partition the dataset into subsets of examples that rely on similar mechanisms. We instantiate this step with clustering over per-example attribution patterns (DCD-Cluster). Across four tasks and three models, DCD-Cluster decomposes the dataset into multiple subsets whose circuits exhibit distinct faithfulness patterns, providing useful hypotheses about mechanistic differences among examples.
| Comments: | 57 pages, 49 figures, 36 tables, Under review |
| Subjects: | Artificial Intelligence (cs.AI) |
| MSC classes: | 68T07 |
| ACM classes: | I.2.7 |
| Cite as: | arXiv:2605.09129 [cs.AI] |
| (or arXiv:2605.09129v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2605.09129 arXiv-issued DOI via DataCite |
Submission history
From: Daking Rai Mr [view email]
[v1]
Sat, 9 May 2026 19:20:57 UTC (1,327 KB)
[v2]
Mon, 12 Oct 2026 13:49:33 UTC (761 KB)
来源：arXiv cs.AI · arxiv.org