@iatitov on Backlist

Variance Reduction for Long-Horizon LLM RL Without a Critic (t.co)

Token-level credit assignment without a value network is a useful step toward cheaper long-horizon training