MaxRL proposes likelihood training for binary-reward RL
Published on February 9, 2026, the post presents MaxRL: additional rollouts approximate maximum-likelihood training with non-differentiable sampling. The post says the method reduces underweighting of difficult prompts and achieves up to 20× greater test-time efficiency than GRPO.
In a post published on February 9, 2026, MaxRL is described as a method for reinforcement learning with binary rewards. It uses additional rollouts to approximate maximum-likelihood training in settings with non-differentiable sampling.
According to the post, the approach reduces underweighting of difficult prompts and achieves up to 20× greater test-time efficiency than GRPO. The text says it may interest engineers assessing reward optimization and compute scalability in RL systems. Consult the original post and verify the experimental conditions and metric behind the comparison.