Salakhutdinov shares TailRL work on arXiv

Russ Salakhutdinov announced Tail-Likelihood Reinforcement Learning, or TailRL. The method extends maximum-likelihood reinforcement learning from binary to continuous rewards. Instead of optimizing only mean reward, TailRL maximizes the expected log of upper-tail probabilities to place more weight on rare high-reward outcomes. The arXiv post notes that two policies can share the same average reward yet differ sharply in their tail behavior. The paper is available at the linked source.