Study Moves From Predicting Model Errors to Auditing Causal Influence of SAE Features
A new paper examines sparse autoencoder features linked to errors in multiple-choice QA, arguing that association alone does not establish causal influence and proposing methods to audit features for causal effects.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-10-30 04:00 · arXiv cs.CL
From Error Prediction to Causal Influence: Auditing SAE Features in Multiple-Choice QA