Base: google/gemma-3-12b-it
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 20) — selected by in-loop MathVista-150 eval_reward
Paper cell: MLLM main table (mmr1, big tier) · gemma-3-12b-it · TTRL
WandB run: https://wandb.ai/logan-yang2002-johns-hopkins-university/co-rl-mllm-mmr1/runs/gemma3_12b_ttrl_mmr1-v1
best repo: q1716523669/mllm-mmr1-ttrl-gemma3-12b · endpoint repo: q1716523669/mllm-mmr1-ttrl-gemma3-12b-endpoint

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

ckpt	MathVision	MathVerse	MathVista	We-Math	avg
best (step 20)	28.36	37.61	37.8	62.93	41.68
endpoint (step 722)	14.67	18.93	32.2	45.29	27.77

best → endpoint: 41.68 → 27.77 — TTRL belief-reinforcement collapse (self-reward saturates while real accuracy degrades).

q1716523669/mllm-mmr1-ttrl-gemma3-12b