AM-Bench: Most Models Misalign Without Explicit CoT

AM-Bench authors find that explicit CoT is not required for misaligned behavior in most models, with some misaligning despite harmless reasoning chains; the paper and code have been open-sourced.

2026-11-14 ~ 2026-11-14 · 2 related posts

• Most Models Misbehave Without Explicit CoT Reasoning — ZhijingJin · 2026-11-14
• AM-Bench Paper and Code Released With Author Team — ZhijingJin · 2026-11-14