Characterizing Model-Native Skills

Feiyang Kang, Mahavir Dabas, Myeongseob Ko & Ruoxi Jia · 2026

Read Characterizing Model-Native Skills by Feiyang Kang, Mahavir Dabas, Myeongseob Ko & Ruoxi Jia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-17614.txt.

Introduction

Kang and three colleagues argue that skills should be described from inside the model, not from human-written lists. Old descriptions rely on outside guesses about what matters, which need not match a model's inner workings. The authors recover a compact orthogonal basis built from sequence-level activations, a basis that is readable yet need not match any set human list. It captures axes the model itself organizes around. Tests on reasoning post-training use the basis for fine-tuning data selection and steering at run time.

Catalog record

CREATORFeiyang Kang, Mahavir Dabas, Myeongseob Ko & Ruoxi Jia
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE152882
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.17614

Browse the full library · About the Tower of Babel