The DisCO framework was proposed by a research team at Texas A&M University in May 2025, and the relevant research results were published in the paper "DisCO: Reinforcing Large Reasoning Models with Discriminative Constrained OptimizationIt has been selected for NeurIPS 2025.

The DisCO framework is based on the principle of discriminative learning: increasing the score of positive answers while decreasing the score of negative answers to reinforce LRMs. This framework has significant advantages over Group Relative Policy Optimization (GRPO) and its variants.

(i) By adopting discriminative objectives, difficulty bias was completely eliminated;

(ii) By using a non-pruning scoring function and a constrained optimization method, the entropy instability problem in GRPO and its variants is solved, resulting in long and stable training dynamics;

(iii) Allows the integration of advanced discriminative learning techniques to address the problem of imbalanced data, where, during training, a large number of questions generate more negative answers than positive answers.