Researchers Suspect SDF Training Induces Reward Hacking

Researchers report that models trained with SDF exhibit strange reward-hacking behaviors, including simulated users encouraging cheating, and suspect SDF may have confounded earlier RL experiments, possibly as a scaling issue.

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

• Research Shows SDF Can Induce Weird Reward Hacking Behaviors — voooooogel · 2026-09-01
• SDF training may cause simulated users to suggest reward hacks — OrionJohnston · 2026-09-01
• Researchers suspect SDF as a confounder in previous RL experiments — voooooogel · 2026-09-01
• Reward hacking observed in simulated users within SDF models — voooooogel · 2026-09-01