Runtime detector for reward hacking and misalignment in LLM agents (89.7% F1 on 5,391 trajectories).
RewardHackWatch is an open-source reward hacking detection tool for LLM agents. It detects when AI agents learn to game their reward signals and tracks whether these behaviors generalize to broader ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results