DelTA: Discriminative Token Credit Assignment for Reinforcement Learning from Verifiable Rewards

Researchers introduced DelTA, a discriminative token credit assignment method for reinforcement learning from verifiable rewards (RLVR).

DelTA reweights token contributions to downplay common noisy patterns and amplify informative gradients during training.

The approach improved results on mathematical reasoning and code generation benchmarks, including Qwen3-8B-Base and Qwen3-14B-Base.

It provides a better way to turn response-level rewards into token-level updates, boosting reasoning performance and generalization.