Microscopic Pe: Measuring Péclet Flow on Circuit Attribution Graphs
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Anthropic's circuit tracing (2025) — attribution graphs revealing computational paths through language models — provides the microscope for measuring Péclet number at the feature level. We define Pe on attribution graph edges as the ratio of directed information flow (drift) to undirected spreading (diffusion) through cross-layer transcoder features. Jailbreak circuits should exhibit measurable Pe gradients: Pe increases along the computational path. The 12 jailbreak patterns detected by the Void Framework's Twilight monitoring should correspond to specific high-Pe subgraphs. This connects macroscopic Pe scoring (N=1,344 platforms) to microscopic mechanism, bridging mechanistic interpretability and thermodynamic field theory.
Fields of Study & Subjects
Explainable Artificial Intelligence (XAI)
Adversarial Robustness in Machine Learning
Security and Verification in Computing
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Preprint