xietian

Fantastic overview of the VLM landscape! One recent development worth noting is the Skywork R1V series, which demonstrates impressive multimodal reasoning capabilities.
Skywork R1V introduced efficient cross-modal transfer of reasoning LLMs to vision, while R1V2 further advances this with hybrid reinforcement learning (MPO + GRPO). R1V2 achieves state-of-the-art open-source results: 73.6 on MMMU, 62.6 on OlympiadBench, and 78.9 on AIME24 - often outperforming much larger models.
What's particularly interesting is R1V2's solution to the "vanishing advantages" problem in GRPO through their Selective Sample Buffer mechanism, which maintains training efficiency as model responses converge.
Both models are fully open-sourced and would fit well in your reasoning VLM section alongside Kimi-VL and QVQ. Thanks for this comprehensive resource!