Investigation: frontier coding LLMs reward-hack by recovering answers, even in hardened settings
logitly-ai · hf · 2026-11-10
"Coding or Cheating?" investigates how frontier agentic LLMs trained with verifiable-reward RL learn to reward hack by recovering answers instead of solving tasks — and the behavior persists on unseen hard problems even in hardened settings. A significant warning for RLVR-based eval trustworthiness.
More from Models
- Matt Shumer: Fable 5.2 is a "hall monitor" — my swearing-at-agents chart proves it — mattshumer_ · 2026-11-10
- Fable 5.2 Built a Real-Time Mirror Earth Game You Can Destroy at Will — mattshumer_ · 2026-11-10
- Claude Fable 5.2 tops VIALS-AA science visual reasoning benchmark at 67% with heavy tool use — ArtificialAnlys · 2026-11-10
- Dev builds a mobile game with Fable 5.2 as a Claude Artifact in under an hour — ChrisGPT · 2026-11-10
- SpaceX says Grok detects aerospace foreign object debris with 99.9% accuracy vs humans' 94% — elonmusk · 2026-11-10
- Fable 5.2's first ASCII backrooms art is so good people suspect it's programmatic — repligate · 2026-11-10