CORTEXA
← Browse

Alice Chan

1 paper indexed

arxivcs.ROcs.AIcs.LGeess.SYmath.OC2026-07-16

Steering Robustness into World Action Models via Mechanistic Interpretability and Optimal Control

Jihoon Hong, Julian Skifstad, Qiyue Dai, Alice Chan, Glen Chou

World Action Models (WAMs) enable semantically- and physically-informed control but are brittle under distribution shift. In this work, we use mechanistic interpretability to study how robustness-relevant perturbations are represented in WAM activation space. Comparing activation…

View free PDFSource page