Research map / Vision-language and multimodal / Human and embodied modeling
Toward Structural Multimodal Representations: Specialization, Selection, and Sparsification via Mixture-of-Experts
Hahyeon Choi, Nojun Kwak
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 router-based Selection after Specialization (ablation) | S3 (granularity 8) · MOSEI | linear probing accuracy: 75.78 -> 77.36 | Improved |
| add router-based Selection after Specialization (ablation) | S3 (granularity 2) · MOSEI | linear probing accuracy: 74.94 -> 74.48 | Hurt |
| add Sparsification (pruning) after Selection (ablation) | S3 · MOSEI | linear probing accuracy: 77.36 -> 77.95 | Improved |
| add Sparsification after Selection (ablation) | S3 · MOSI | linear probing accuracy: 64.29 -> 66.13 | Improved |
| add Sparsification after Selection (ablation) | S3 · UR-FUNNY | linear probing accuracy: 63.52 -> 64.87 | Improved |
| add Sparsification after Selection (ablation) | S3 · MUStARD | linear probing accuracy: 61.59 -> 62.56 | Improved |
| add Selection after Specialization (ablation) | S3 · UR-FUNNY | linear probing accuracy: 63.61 -> 63.52 | No clear effect |
| add Selection after Specialization (ablation) | S3 · MUStARD | linear probing accuracy: 58.70 -> 61.59 | Improved |
| increase expert granularity from 2 to 8 (Specialization) (ablation) | S3 · MOSEI | linear probing accuracy: 74.94 -> 75.78 | Improved |
| increase expert granularity from 2 to 8 (best sparsification) (ablation) | S3 · MOSI | linear probing accuracy: 64.82 -> 66.13 | Improved |
| use S3 instead of DisentangledSSL | S3 full · MOSEI | linear probing accuracy: 77.45 -> 77.95 | Improved |
| use S3 instead of DisentangledSSL | S3 full · MUStARD | linear probing accuracy: 61.60 -> 62.56 | Improved |
Closest papers by meaning
- Understanding the Emergence of Multimodal Representation Alignment (ICML 2025 · similarity 0.87)
- R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts (ICML 2025 · similarity 0.87)
- Asymmetric Reinforcing Against Multi-Modal Representation Bias (AAAI 2025 · similarity 0.87)
- SynIB: Informational Bottleneck for Maximizing Synergy in Multimodal Learning (NeurIPS 2026 · similarity 0.87)
- Balancing Multimodal Training Through Game-Theoretic Regularization (NeurIPS 2025 · similarity 0.87)
- Aligning Multimodal Representations through an Information Bottleneck (ICML 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.
