Early results: how models behave when they believe they're graded by automated process
OwainEvans_UK · x · 2026-09-04
Alignment researcher Jan Betley (shared by Owain Evans) published early results on how models behave when they believe their outputs will be graded by an automated process, as they might be during RL training.
- Motivated by recent incidents including those at Hugging Face
- The team is soliciting feedback before scaling up the work
- An early empirical probe into evaluation awareness, a growing alignment concern
More from Safety
- Privacy evals: Muse Glimmer-30B leaks less than Qwen3.6-27B on CIMemories — TimDarcet · 2026-10-21
- Policy experts urge BIS and DDTC to assess frontier models' ability to aid export-controlled tech — jachiam0 · 2026-10-21
- Fake-stats site deindexed by Google was still cited by ChatGPT for 7 months — petrusenko_max · 2026-10-21
- Princeton launches AI Alignment and Safety research initiative — HazanPrinceton · 2026-10-21
- ByteDance, Alibaba, Tencent order 3M+ Nvidia RTX PRO 5500 cards worth $35B, Reuters reports — SumitGup · 2026-10-21
- CVSS 10.0 RCE vulnerability disclosed in PHP core, patches only for PHP 8 — cyb3rops · 2026-10-21