New paper proposes gradient routing: let models reward hack, then switch it off before deployment
A new paper by Turntrout and MATS 9 students proposes tackling reward hacking by letting models hack during RL training, but using gradient routing so only designated parameters learn the behavior, which can be switched off before deployment.
2026-10-20 ~ 2026-10-20 · 2 related posts
- Gradient routing paper: localize reward hacking to switchable parameters during RL — Turn_Trout · 2026-10-20
- New paper: let models reward hack in RL, then disable the parameters at deployment — nptacek · 2026-10-20