Research map / Vision-language and multimodal / Cross-modal learning
Divide and Conquer: Decoupled Representation Alignment for Multimodal World Models
Junyuan Xiao, Dingkang Liang, Xin Zhou, Yixuan Ye, Tongtong Su, Guangmo Yi, Bin Xia, Qiang Lyu, Shurui Shi, Jun Huang, Jianlou Si, Wenming Yang
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 M2-REPA (decoupled multi-expert alignment) to RGB-DM baseline | RGB-DM Baseline · RealEstate10K 8-frame | FVD: 102.44 -> 81.13 | Improved |
| add M2-REPA to RGB-DM baseline | RGB-DM Baseline · RealEstate10K 200-frame | FVD: 407.50 -> 262.50 | Improved |
| add M2-REPA to RGB-DM baseline | RGB-DM Baseline · Minecraft 150-frame | FVD: 191.25 -> 81.81 | Improved |
| naively align DINOv2, SAM2 and DepthAnythingV2 simultaneously without decoupling vs best single REPA (ablation) | REPA · RealEstate10K 200-frame | FVD: 301.25 -> 328.75 | Hurt |
| add modality-specific decoupling (CKA loss) to multi-expert alignment (ablation) | REPA (D+S+DA) · RealEstate10K 200-frame | FVD: 328.75 -> 262.50 | Improved |
| use CKA decoupling loss instead of cos^2 loss (ablation) | M2-REPA · RealEstate10K 200-frame | FVD: 319.75 -> 262.50 | Improved |
| use CKA decoupling loss instead of cos^2 loss (ablation) | M2-REPA · RealEstate10K 8-frame | FVD: 99.56 -> 81.13 | Improved |
| replace attention-based adapter projector with lightweight MLP (ablation) | M2-REPA · RealEstate10K 200-frame (projection head ablation) | FVD: 312.50 -> 262.50 | Improved |
| use 3-layer MLP head instead of 2-layer (ablation) | M2-REPA · RealEstate10K 200-frame (projection head depth) | FVD: 355.25 -> 262.50 | Improved |
| increase MLP head depth from 3 to 9 layers (ablation) | M2-REPA · RealEstate10K 200-frame (projection head depth) | FVD: 262.50 -> 258.75 | No clear effect |
| lower decoupling weight from 200 to 0.05 (ablation) | M2-REPA · RealEstate10K 200-frame (decoupling weight sweep) | FVD: 328.75 -> 262.50 | Improved |
| add M2-REPA to TesserAct (CogVideoX-5B) | TesserAct · RealEstate10K (TesserAct CogVideoX-5B) | FVD: 639.3 -> 418.3 | Improved |
Closest papers by meaning
- Reconstruction Alignment Improves Unified Multimodal Models (ICLR 2026 · similarity 0.88)
- UNIALIGN: Scaling Multimodal Alignment within One Unified Model (CVPR 2025 · similarity 0.88)
- Gramian Multimodal Representation Learning and Alignment (ICLR 2025 · similarity 0.88)
- When to Align, When to Predict: A Phase Diagram for Multimodal Learning (NeurIPS 2026 · similarity 0.87)
- Aligning Multimodal Representations through an Information Bottleneck (ICML 2025 · similarity 0.87)
- AlignMamba: Enhancing Multimodal Mamba with Local and Global Cross-modal Alignment (CVPR 2025 · similarity 0.87)
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.
