Research map / Vision-language and multimodal / Multimodal LLMs
LaViT: Aligning Latent Visual Thoughts for Multi-modal Reasoning
Linquan Wu, Tianxiang Jiang, Yifei Dong, Haoyu Yang, Fengji Zhang, Shichang Meng, AI Xuan, Linqi Song, Jacky Keung
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory alignment loss (ablation) | LaViT · MMVP | accuracy: 64.33 -> 67.33 | Improved |
| add semantic reconstruction loss (ablation) | LaViT · MMVP | accuracy: 65.33 -> 67.33 | Improved |
| add semantic reconstruction loss (ablation) | LaViT · BLINK Relative Reflectance | accuracy: 42.54 -> 45.52 | Improved |
| use curriculum sensory gating instead of hard switch (ablation) | LaViT · MMVP | accuracy: 59.33 -> 67.33 | Improved |
| use curriculum sensory gating instead of hard switch (ablation) | LaViT · BLINK Relative Reflectance | accuracy: 48.51 -> 45.52 | Hurt |
| use generated latent tokens at inference (ablation) | LaViT · BLINK IQ-Test | accuracy: 25.33 -> 32.0 | Improved |
| use two-stage gated curriculum instead of single-stage training (ablation) | LaViT · BLINK Relative Reflectance | accuracy: 38.81 -> 45.52 | Improved |
| use 4 instead of 6 or 8 latent tokens (ablation) | LaViT · MMVP | accuracy: 61.33 -> 67.33 | Improved |
| train with latent visual thought alignment instead of text-only SFT | LaViT · BLINK Relative Reflectance | accuracy: 34.33 -> 45.52 | Improved |
| apply LaViT to Qwen2.5-VL-3B | LaViT · BLINK Relative Depth | accuracy: 61.29 -> 78.23 | Improved |
| compare 3B LaViT with 7B LVR | LaViT · BLINK Spatial Relation | accuracy: 89.5 -> 81.82 | Hurt |
| distill teacher attention into latent tokens | LaViT · BLINK Relative Depth attention entropy | mean entropy: 4.870 -> 4.686 | Improved |
Closest papers by meaning
- Leveraging Latent Visual Reasoning in Silence (NeurIPS 2026 · similarity 0.92)
- Latent Visual Reasoning (ICLR 2026 · similarity 0.92)
- Monet: Reasoning in Latent Visual Space Beyond Image and Language (CVPR 2026 · similarity 0.91)
- UniVLR: Unifying Text and Vision in Visual Latent Reasoning for Multimodal LLMs (NeurIPS 2026 · similarity 0.91)
- Visual Enhanced Depth Scaling for Multimodal Latent Reasoning (NeurIPS 2026 · similarity 0.91)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
