List your company

Multimodal Math Reasoning on R1-OV Bench

38.9AccuracyAVATAR
34.53235.66636.837.934Aug 5, 2025

Evaluation results

MethodLinks
AVATARbackbone=Qwen2.5-VL-7B2025.0838.9
AVATARbackbone=Qwen2.5-VL-7B, ablation=w/o Replay Buffer (on-policy)2025.0838.5
VLAAThinker-VL-7B2025.0838.4
AVATARbackbone=Qwen2.5-VL-7B, ablation=w/o TAS (uniform credit)2025.0837.5
AVATARbackbone=Qwen2.5-VL-7B, ablation=w/o Hinting2025.0837
Qwen2.5-VL-7B + GRPOtraining=GRPO2025.0836.8
Qwen2.5-VL-7Bstatus=Baseline2025.0834.9
OpenVLThinker-7B2025.0834.7