Cross-family co-learning (co-GRPO-DP): trained jointly with peer Qwen/Qwen2.5-VL-7B-Instruct via majority-vote cross-labeling — no ground-truth labels are used.

Base: OpenGVLab/InternVL3_5-8B-HF · Peer: Qwen/Qwen2.5-VL-7B-Instruct
Method: co-GRPO-DP (heterogeneous co-learning, rendezvous cross-pseudo-labels)
Dataset: mmr1 (~8k), 1 epoch = 722 steps
Checkpoint: best (best_model)
4-bench avg: **54.01 ** (MathVision / MathVerse / MathVista / We-Math; prompt=answer, greedy T=0, mathruler)
WandB project: co-rl-mllm-mmr1