Interpretability without actionability: mechanistic methods cannot correct language model errors despite near-perfect internal representations

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Basu, Sanjay, Patel, Sadiq Y., Sheth, Parth, Muralidharan, Bhairavi, Elamaran, Namrata, Kinra, Aakriti, Morgan, John, Batniji, Rajaie
Format:	Preprint
Published:	2026
Subjects:	Artificial Intelligence I.2.7; J.3
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Internet

https://arxiv.org/abs/2603.18353