Reward Hacking MO Checkpoints and Rollouts - a ai-safety-institute Collection

ai-safety-institute 's Collections

Reward Hacking MO Checkpoints and Rollouts

RealityTest benchmark

(Some) Emergent Misalignment from Reward Hacking in RL

Were You Honest Probes

Lie Detection Model Organisms

Were You Truthful Probes

Targeted Apollo Deception Probes

Did You Lie Probes

Catch a Liar: Unrelated Questions Classifier

Apollo-Style Deception Probes

Lie Detection Model Organisms Datasets

Lie Detection Model Organisms Merged

Gender Secret Hyperparameter Sweep

Reward Hacking MO Checkpoints and Rollouts

updated 14 days ago

Artefacts from "(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL". KL 0.0 hacks with faithful CoT, 0.2 with unfaithful.