From P(y|x) to P(y): Investigating Reinforcement Learning in Pre-train Space

Open paper
Benchmark Model Rank Results
code-generation-on-humanevalDSRL (Qwen3-8B)#28Pass@1: 90.24
mathematical-reasoning-on-minerva-mathDSRL (Qwen3-8B)#37Accuracy: 30.32
mathematical-reasoning-on-olympiadbenchDSRL (Qwen3-8B)#41Accuracy: 41.82
mathematical-reasoning-on-math-500DSRL (Qwen3-8B)#58Accuracy: 90.31
mathematical-reasoning-on-aime24DSRL (Qwen3-8B)#74Acc: 56.15
multi-task-language-understanding-on-mmlu-proDSRL (Qwen3-8B)#86Accuracy: 63.79
mathematical-reasoning-on-aime-2025DSRL (Qwen3-8B)#93Accuracy: 42.19
multiple-choice-question-answering-mcqa-on-gpqa-diamondDSRL (Qwen3-8B)#109Accuracy: 53.48

Rank counts only results with a code link.