Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision

Benchmark Model Rank Results
math-word-problem-solving-on-mathSD-Zero (Olmo-3-7B-Instruct)#2Accuracy: 94.0
mathematical-reasoning-on-amo-benchSD-Zero (Qwen3-4B-Instruct)#5Accuracy: 16.0
mathematical-reasoning-on-hmmt-feb-2025SD-Zero (Qwen3-4B-Instruct)#16Accuracy: 45.4
mathematical-reasoning-on-aime24SD-Zero (Qwen3-4B-Instruct)#53Acc: 68.3
mathematical-reasoning-on-aime-2025SD-Zero (Qwen3-4B-Instruct)#72Accuracy: 60.0