Gaussian GRPO-trained multimodal reasoner achieving strong InfoVQA. Uses distributional matching for inter-task gradient equity across diverse visual tasks.

Introduces Gaussian GRPO (G2RPO), replacing standard linear scaling in GRPO with non-linear distributional matching. OpenVLThinkerV2-7B achieves new SOTA for open-source 7B models on MMMU (71.6%), Mat

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

8B open-weight multimodal model trained with GRPO+GDPO. Competitive with Gemini 2.5 Pro on DocVQA and chart understanding. New SOTA for open-weight VLMs on MMMU.

+1.6 over Qwen2.5-VL-3B baseline. Spatial priors help on info-heavy documents.

Layout-aware reasoning with Visual-Semantic Chain. Achieves 78.6 on InfoVQA, outperforming Qwen3-VL-8B by 2.9 points.

+2.0 over Qwen3-VL-2B baseline. Small model benefits most from spatial priors.

4B end-to-end OCR model that ranks #1 on OmniDocBench among end-to-end models. Introduces "Layout-as-Thought" for structured layout representations. Outperforms Qwen3-VL-4B on ChartQA (+4.8) and Chart