AutoCircuit: A New Framework for Automatically Discovering Interpretable Reasoning Circuits in Large Language Models

The AutoCircuit project from AI Safety Camp 2025 proposes a systematic method for automatically discovering interpretable reasoning circuits inside Transformer models. By mining attribution graphs and combining analysis with LLM agents, it aims to lower the barrier to mechanistic interpretability research and enable real-time safety monitoring. The project's core goal is to systematically identify stable computational circuits in models to support AI safety and alignment research.

As LLM capabilities rapidly improve, understanding their internal working mechanisms becomes increasingly important. Anthropic's 2025 attribution graph method opened a new path for mechanistic interpretability, but manually analyzing large numbers of graphs to identify common computational patterns is impractical. As Project No.24 of AI Safety Camp 2025, AutoCircuit's core goal is to mine attribution graphs generated by Neuronpedia via data mining, use LLM agents to analyze graphs across prompt categories, and identify stable reasoning circuits.

AutoCircuit adopts a four-stage technical architecture:

1. Automated Graph Collection: Use Neuronpedia API to batch generate attribution graphs for different prompt categories (fact recall, arithmetic operations, etc.) to improve coverage and efficiency;
2. Graph Simplification Algorithm: Filter noise nodes and retain core computational structures;
3. Pattern Recognition: Analyze cross-context graphs to identify repeatedly occurring circuit motifs;
4. Causal Validation: Validate the causal role of circuits through interventions like feature ablation and activation patching.

AutoCircuit integrates existing interpretability infrastructure: it uses Anthropic 2025 cross-layer transcoders and attribution graph construction algorithms, combined with Neuronpedia's model steering API; it leverages Claude Sonnet as an agent to analyze adjacency matrix patterns in graphs, propose circuit hypotheses, and explain activation co-occurrences; it develops quantitative metrics such as graph completeness scores and indirect influence matrix analysis to guide hypothesis optimization and support manual verification.

To address the false positive risk of automated circuit discovery, the project designs a multi-layer validation mechanism: it requires multiple independent confirmation signals to accept a circuit hypothesis, and key safety findings undergo human-in-the-loop validation; if automated annotation is unreliable, it switches to semi-automation (AI proposes explanations + manual verification); it uses graph structure metrics (centrality, node distance, etc.) to filter circuit subsets, which are manually confirmed by researchers to control agent bias.

AutoCircuit is expected to produce a curated library of interpretable reasoning circuits (with evidence of causal effects), which is of great significance for AI safety:

1. Democratize mechanistic interpretability research and lower professional barriers;
2. Support real-time safety monitoring and proactively identify signs of model misalignment;
3. Accelerate AI alignment research and enable targeted interventions in model decision-making processes.

The project progresses in three phases: Phase 1 implements automated circuit discovery and feature annotation; Phase 2 conducts systematic validation and exploration; Phase3 develops cross-model analysis and deployment frameworks. Deliverables include: a circuit library published on Neuronpedia, open-source code on GitHub, an arXiv paper, and conference submissions. The minimum goal is a semi-automated research accelerator, and the vision is a fully automated interpretability platform that monitors dangerous capabilities in real time and provides intervention measures.