Researchers probed whether LLMs internally represent all 26 CATO legal factors. They trained linear probes to read factor presence from model activations. They also tested sparse autoencoder features for localized readouts. All 26 factors proved linearly decodable across three open models. Single-feature readouts achieved AUC scores from 0.86 to 1.00. Five selected features matched full-vector performance. The team reported 78 model-specific layer–feature coordinates. A controlled corpus of 6,992 passages validated the labels. Human audits confirmed the factor assignments. Code and corpus are publicly available.
What it unlocks
Reading expert-defined legal factors directly from LLM internal activations and sparse autoencoder features.
- 26 CATO factors
- 3 open models tested
- 78 layer–feature coordinates reported
- 0.86 to 1.00 best-layer test AUC range
What you need to act on it
- controlled natural-language passages
- publicly released SAEs for Qwen3 and Gemma-2
- CATO factor definitions
Sources