pith. sign in

← back to paper

Review history

arxiv: 2604.23130 · 2 revisions

From Concept-Aligned Tokens to Vulnerable Features: Mechanistic Localization of Jailbreaks

  1. 2026-07-04 CONDITIONAL MODERATE v1.1.0-glm novelty 4.0
    157878 ms 10897 in 3470 out 2026-07-04T15:41:22.433513+00:00
  2. 2026-05-08 UNVERDICTED LOW v0.9.0 novelty 4.0
    50879 ms 5517 in 1397 out 2026-05-08T08:23:27.811045+00:00