Base: OpenGVLab/InternVL3_5-8B
Method: TTRL — self-labeling GRPO (majority-vote pseudo-labels, no GT)
Dataset: mmr1 (~8k), 1 epoch = 722 steps
This checkpoint: BEST (best-by-val, step 460) — selected by in-loop MathVista-150 eval_reward
Paper cell: MLLM main table (mmr1, big tier) · InternVL3_5-8B · TTRL
WandB run: https://wandb.ai/logan-yang2002-johns-hopkins-university/co-rl-mllm-mmr1/runs/internvl35_8b_ttrl_mmr1-v1
best repo: q1716523669/mllm-mmr1-ttrl-internvl35-8b · endpoint repo: q1716523669/mllm-mmr1-ttrl-internvl35-8b-endpoint

Eval — 4 benchmarks (prompt=answer · greedy T=0 · mathruler)

ckpt	MathVision	MathVerse	MathVista	We-Math	avg
best (step 460)	38.03	43.1	28.2	74.25	45.9
endpoint (step 722)	38.82	43.43	28.7	73.62	46.14

q1716523669/mllm-mmr1-ttrl-internvl35-8b