TY  - RPRT
TI  - Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations
AU  - Sanjay Basu
AU  - Sadiq Y. Patel
AU  - Parth Sheth
AU  - Bhairavi Muralidharan
AU  - Namrata Elamaran
AU  - Aakriti Kinra
AU  - John Morgan
AU  - Rajaie Batniji
PY  - 2026
UR  - https://arxiv.org/abs/2603.18353
ID  - 2603.18353
ER  -