arXiv cs.CL· Shariar Kabir, Hong Kiat Tan, Swastik Agrawal, Sai V R Chereddy, Sriram Balasubramanian·· 11 小时前AI 评分31
Circuit Oracle：用自然语言查询自动分析归因图
Circuit Oracle: Automating Attribution Graph Analysis via Natural-Language Queries
AI 导读
Circuit Oracle 是一个配备多智能体框架和任务专用技能的现成 LLM，通过工具调用读取归因图来回答关于目标模型的自然语言问题。在虚假特征检测任务上达到 78.2 ± 2.4% 准确率，优于单层特征排序基线；但在秘密引出任务上不及 Activation Oracles，在 50 条提示的拒绝抑制任务上单次干预略低于 diff-in-means 基线。
正文 · 原文
Abstract:Attribution graphs, an emerging tool in mechanistic interpretability, use transcoders to decompose language model computations into sparse interpretable features connected by causal edges. However, turning a graph into a safety-relevant insight requires hours of manual analysis by experts. We introduce Circuit Oracle, an off-the-shelf LLM equipped with a multi-agent harness and task-specific authored skills, which answers natural-language questions about a target model by reading the attribution graph through tool calls. We evaluate it on three safety-relevant proxy tasks: detecting spurious features in probe circuits, eliciting hidden knowledge from taboo-finetuned models, and suppression jailbreaking via feature-level interventions. Beyond task scores, the oracle produces pruned subgraphs and context-specific feature explanations that we assess qualitatively. Results are mixed. On spurious-feature detection, Circuit Oracle reaches 78.2 $\pm$ 2.4% accuracy, outperforming single-layer feature-ranking baselines. On secret elicitation, trained self-explainers like Activation Oracles remain the stronger baseline. On refusal suppression over 50 prompts, its single committed intervention is slightly below a diff-in-means baseline, and only exceeds it when allowed the best of five attempts. Results are stable across five runs and across model choices for the agents. We also release the pruned subgraphs and context-specific feature readings that revise existing global labels, and invite future work to develop evaluations that can measure their usefulness directly.
| Comments: | Extended version of a paper presented at the Mechanistic Interpretability and Compositional Learning workshops at ICML 2026 |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.00420 [cs.CL] |
| (or arXiv:2611.00420v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.00420 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Shariar Kabir [view email]
[v1]
Fri, 30 Oct 2026 22:12:53 UTC (1,335 KB)
来源：arXiv cs.CL · arxiv.org