Dynamic Important Example Mining for Reinforcement Finetuning

Benchmark Model Rank Results
mathematical-reasoning-on-mathverseQwen2.5-VL-32B + DIEM#13Accuracy: 58
multimodal-reasoning-on-mmmuQwen2.5-VL-32B + DIEM#23Accuracy: 71.9
vision-language-understanding-on-mmstarQwen2.5-VL-32B + DIEM#37Accuracy: 71.2
multimodal-reasoning-on-mathvisionQwen2.5-VL-32B + DIEM#39Accuracy: 40.3
multimodal-reasoning-on-mathvistaQwen2.5-VL-32B + DIEM#40Accuracy: 76.9
mathematical-reasoning-on-math-500Qwen3-4B + DIEM#115Accuracy: 65.9
mathematical-reasoning-on-aime24Qwen3-4B + DIEM#128Acc: 16.2
mathematical-reasoning-on-aime-2025Qwen3-4B + DIEM#138Accuracy: 14.2