We introduce Reasoning via Video, a new paradigm that uses maze-solving video generation to probe multimodal reasoning; our VR-Bench shows that fine-tuned video models consistently outperform strong VLMs on long-horizon spatial planning tasks.
我的Fcitx5-rime配置：雾凇拼音+万象语法大模型
自动同步雾凇拼音和万象拼音的 Rime 输入法配置，每日更新词库和语言模型
Evaluate vision-language model perception by isolating failures in visual encoders and language heads using standardized tasks and probes.
🖥️ Install and configure Fcitx5 with Rime for an enhanced input experience, featuring powerful grammar models and a rich vocabulary.