Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL

Benchmark Model Rank Results
mathematical-reasoning-on-mathverseCo-RL (InternVL-3.5-8B, Different family)#27Accuracy: 40.74
multimodal-reasoning-on-mathvisionCo-RL (InternVL-3.5-8B, Different family)#45Accuracy: 35.3
multimodal-reasoning-on-mathvistaCo-RL (Qwen2.5-VL-7B, Different family)#61Accuracy: 71.0
code-generation-on-mbppCo-RL (Qwen2.5-7B, Different family)#71Accuracy: 65.6
code-generation-on-livecodebench-v6Co-RL (Qwen2.5-7B, Different family+)#76Pass@1: 0.266
code-generation-on-humanevalCo-RL (Llama-3.1-8B-Instruct, Same family)#86Pass@1: 68.3
arithmetic-reasoning-on-gsm8kCo-RL (Llama-3.1-8B-Instruct, Different family+)#107Accuracy: 85.4
mathematical-reasoning-on-math-500Co-RL (Qwen2.5-7B, Different family)#109Accuracy: 75.2
multiple-choice-question-answering-mcqa-on-gpqa-diamondCo-RL (Qwen2.5-7B, Different family+)#129Accuracy: 37.9