Research map / Vision-language and multimodal / Multimodal LLMs
Embed-RL: Reinforcement Learning for Reasoning-Driven Multimodal Embeddings
Haonan Jiang, Yuji Wang, Yongjie Zhu, Xin Lu, Wenyu Qin, Meng Wang, Pengfei Wan, Yansong Tang
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 embedder-guided RL (EG-RL) stage (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 65.3 -> 66.8 | Improved |
| add weighted negative sampling to outcome reward (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 66.5 -> 66.8 | Improved |
| add process reward (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 66.0 -> 66.8 | Improved |
| add outcome reward (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 65.8 -> 66.8 | Improved |
| add reasoning process to T-CoT (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 65.5 -> 66.8 | Improved |
| add reasoning process to T-CoT (ablation) | Embed-RL-2B · MMEB-V2 video overall | score: 50.5 -> 52.1 | Improved |
| add multimodal cues (bounding boxes, keyframes) to T-CoT (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 65.8 -> 66.8 | Improved |
| add T-CoT instead of raw input (ablation) | Embed-RL-2B · MMEB-V2 (Embed-RL-2B) | overall score: 60.2 -> 66.8 | Improved |
| add T-CoT instead of raw input (ablation) | Embed-RL-2B · MMEB-V2 video overall | score: 43.7 -> 52.1 | Improved |
| add T-CoT instead of raw input (ablation) | Embed-RL-2B · MMEB-V2 visdoc overall | score: 72.4 -> 74.1 | Improved |
| scale embedder from 2B to 4B | Embed-RL · MMEB-V2 (All) | score: 66.8 -> 68.1 | Improved |
| Embed-RL-4B vs UME-R1-7B | Embed-RL-4B · MMEB-V2 (All) | score: 64.5 -> 68.1 | Improved |
Closest papers by meaning
- UME-R1: Exploring Reasoning-Driven Generative Multimodal Embeddings (ICLR 2026 · similarity 0.94)
- Think Then Embed: Generative Context Improves Multimodal Embedding (ICLR 2026 · similarity 0.93)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Point-RFT: Improving Multimodal Reasoning with Visually Grounded Reinforcement Finetuning (NeurIPS 2025 · similarity 0.89)
- Revisual-R1: Advancing Multimodal Reasoning From Optimized Cold Start to Staged Reinforcement Learning (ICLR 2026 · similarity 0.89)
- GThinker: Towards General Multimodal Reasoning via Cue-Guided Rethinking (CVPR 2026 · similarity 0.88)
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.
