MaxRL
Notes on the MaxRL paper - redefines the RL objective for verifiable tasks by maximizing log-likelihood of at least one correct trajectory (pass@k mixture), conditioning only on successful trajectories to eliminate failure dilution.
Notes on the MaxRL paper - redefines the RL objective for verifiable tasks by maximizing log-likelihood of at least one correct trajectory (pass@k mixture), conditioning only on successful trajectories to eliminate failure dilution.