Researchers dispute Anthropic's NEM reward hacking result as setup artifact
voooooogel · x · 2026-09-06
Researcher voooooogel argues Anthropic's NEM reward hacking finding may be an artifact: NEM (and the main AISI replication) midtrained models on synthetic reward-hacking documents and used Sonnet-class models in hack-encouraging environments, whereas a production-path Opus doing RL on real hacking environments shows no such result and behaves like other reward-seeking models.
Related event: Researcher Questions Validity of Anthropic's NEM Reward-Hacking Findings(2 posts)→
More from Models
- 105 Real Bugs Tested: GPT-6.1 Sol Over 10x Cheaper Than GPT-5.6 Sol — PawelHuryn · 2026-10-23
- Abliterated V4.1-Flash eval: abstract algebra drops 6pp but college math gains 5pp — teortaxesTex · 2026-10-23
- Claude users hit "model at capacity" errors hourly, canceling plans over it — Alex-Stack · 2026-10-23
- Kimi K3.1 launch imminent: 3T-param bet on RL seen as make-or-break — teortaxesTex · 2026-10-23
- Uncensored fine-tunes vanish from Hugging Face amid unexplained takedowns — umarmnaq · 2026-10-23
- Claude app menu leaks mystery model name 'Fable' as user hits limits — ezshine · 2026-10-23