The latest Data Phoenix news tagged reward-hacking.
Apple researchers derived an exact reward-gain expression for an optimal KL-regularized alignment policy and found best-of-N closer to their…