ICML 2026: First Parallel Thinking Framework for Vision-Language Models
ICML 2026|首个视觉语言模型并行思考框架,一文解析内在机制
Visual Para-Thinker introduces a parallel thinking framework for vision-language models, using Pa-Attention and LPRoPE to isolate four visual reasoning paths and training on 163,000 question-answer pairs.
Why it matters: Visual Para-Thinker, accepted at ICML 2026, splits visual reasoning into four parallel paths with Pa-Attention and LPRoPE, trained on 163,000 question-answer pairs.