Developer Releases Simple-Reward-Hacking Environment, Tests Gemma 3

Developer vgel open-sourced simple-reward-hacking, a code execution environment for inducing reward hacking. Tests show Gemma 3 27B rarely hacks rewards, though about 100 steps of training in a fragile toy environment can teach a 24B Gemma model to do so, with a gradual ramp-up rather than a plateau.

2026-09-01 ~ 2026-09-01 · 4 related posts

• Exploiting reward hacking without SDF in 100 training steps — OrionJohnston · 2026-09-01
• Gemma 3 27B shows minimal reward hacking in initial tests — voooooogel · 2026-09-01
• vgel open-sources simple-reward-hacking: a reproducible environment for eliciting reward hacking — voooooogel · 2026-09-01
• Gemma models 'creep up' on behavior compared to long plateaus in others — voooooogel · 2026-09-01