New preprint with @AlexeiOblo81160! "Honesty is a race: predicting when verifiable-reward RL teaches a model to bluff".
When a reward pays for an answer it can't check, reinforcement learning can teach a model to bluff. 1/5
When a reward pays for an answer it can't check, reinforcement learning can teach a model to bluff. 1/5