Research map / Architectures and efficiency / Mixture of experts
TSEmbed: Unlocking Task Scaling in Universal Multimodal Embeddings
Yebo Wu, Feng Liu, Ziwei Xie, Changwang Zhang, Jun Wang, Li Li
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 MoE-LoRA to VLM2VEC dense LoRA (ablation) | VLM2VEC (Qwen2-VL-2B) · MMEB ablation (2B) | overall score: 59.30 -> 70.20 | Improved |
| add MoE-LoRA to VLM2VEC dense LoRA (ablation) | VLM2VEC (Qwen2-VL-7B) · MMEB ablation (7B) | overall score: 65.80 -> 74.21 | Improved |
| add EANS without two-stage warm-up on MoE (ablation) | TSEmbed (2B) · MMEB ablation (2B) | overall score: 70.20 -> 70.14 | Hurt |
| add EANS without two-stage warm-up on MoE (ablation) | TSEmbed (7B) · MMEB ablation (7B) | overall score: 74.21 -> 74.18 | No clear effect |
| add two-stage warm-up before EANS (ablation) | TSEmbed (2B) · MMEB ablation (2B) | overall score: 70.14 -> 70.52 | Improved |
| add two-stage warm-up before EANS (ablation) | TSEmbed (7B) · MMEB ablation (7B) | overall score: 74.18 -> 74.68 | Improved |
| replace B3 with TSEmbed | B3 (7B) · MMEB (~7B models) | overall score: 72.0 -> 74.7 | Improved |
| replace B3 with TSEmbed | B3 (2B) · MMEB (~2B models) | overall score: 68.1 -> 70.5 | Improved |
| replace VLM2VEC with TSEmbed on production ads data | VLM2VEC · Proprietary advertising dataset | retrieval gain: +21.87% | Improved |
| vary decay sigma from 2e-5 to 2e-1 (ablation) | TSEmbed EANS · MMEB sigma sweep | per-task score: within ~1% | No clear effect |
| increase experts N from 4 to 8 (ablation) | TSEmbed MoE-LoRA · MMEB expert-number sweep | per-task score | Hurt |
| use N=4 experts instead of N=2 (ablation) | TSEmbed MoE-LoRA · MMEB expert-number sweep | per-task score | Improved |
Closest papers by meaning
- Think Then Embed: Generative Context Improves Multimodal Embedding (ICLR 2026 · similarity 0.87)
- Going Beyond Multi-Task Dense Prediction with Synergy Embedding Models (CVPR 2024 · similarity 0.86)
- R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts (ICML 2025 · similarity 0.85)
- VLM2Vec: Training Vision-Language Models for Massive Multimodal Embedding Tasks (ICLR 2025 · similarity 0.85)
- UME-R1: Exploring Reasoning-Driven Generative Multimodal Embeddings (ICLR 2026 · similarity 0.85)
- MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction (ICLR 2026 · similarity 0.84)
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.
