视觉推理「囫囵吞枣」靠不住！ProLaViT教会多模态大模型在隐空间里「步步为营」

TL;DR - ProLaViT is an ECCV 2026-accepted framework that improves multimodal visual reasoning through progressive latent-space steps rather than costly image generation or brittle one-shot prediction. It reports a 75.11% average accuracy across several visual-reasoning benchmarks.

• Structures visual reasoning as “Locate → Focus → Isolate” or “Hypothesize → Critique → Verify” chains.
• Uses endogenous self-distillation from the model’s frozen visual encoder and programmatically generated intermediate images, avoiding external vision experts.
• Introduces distance-weighted diversity loss to prevent latent-step representations from collapsing while preserving continuity between adjacent steps.
• Built on Qwen2.5-VL-7B-Instruct, it particularly improves complex tasks such as VisPuzzle and BLINK-Jigsaw.