Research map / Vision-language and multimodal / Cross-modal learning
Training-Free Multimodal Large Language Model Orchestration
Tianyu Xie, Yuexiao Ma, Yuhang Wu, Wang Chen, Jiayi Ji, Tat-Seng Chua, Xiawu Zheng, Rongrong Ji
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 orchestration over Qwen2-VL on long videos | Qwen2-VL · Video-MME (long) | accuracy: 56.3 -> 69.2 | Improved |
| add orchestration over Qwen2-VL overall | Qwen2-VL · Video-MME (overall) | accuracy: 64.9 -> 71.9 | Improved |
| add orchestration over LLaVA-Video on long videos | LLaVA-Video · Video-MME (long) | accuracy: 58.7 -> 72.6 | Improved |
| add orchestration over LLaVA-Video overall | LLaVA-Video · Video-MME (overall) | accuracy: 68.6 -> 76.4 | Improved |
| add orchestration over Qwen2-VL on short videos | Qwen2-VL · Video-MME (short) | accuracy: 75.0 -> 76.7 | Improved |
| upgrade expert pair from Qwen2.5-VL-32B/72B to Qwen3-VL-32B/235B-A22B (ablation) | orchestration with fixed controller · Video-MME | accuracy: 65.58 -> 72.31 | Improved |
| enable evidence-keyed cross-modal memory (ablation) | orchestration framework · dialogue sessions | expert calls: 49% reduction | Improved |
| replace the controller under a fixed token protocol (ablation) | orchestration framework · Video-MME | accuracy | No clear effect |
| use training-free controller instead of supervised LoRA-trained router (ablation) | routing · Video-MME + WorldSense | valid control-token success (%): 99.20 -> 99.10 | No clear effect |
| increase average active expert params through routing from 19.7B to 53.8B (ablation) | Ours (Qwen2.5-VL) · Video-MME | accuracy: 62.07 -> 65.58 | Improved |
| orchestrate experts instead of dense 72B model | Ours vs Qwen2.5-VL-72B · Video-MME | accuracy: 65.74 vs 65.58 | No clear effect |
| memory reuse in round 2 of three-round dialogue | Ours · Video-MME interactive | accuracy: 66.70 -> 69.20 -> 64.30 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VITA1.5 — VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.88)
- VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction (NeurIPS 2025 · similarity 0.88)
- MSP: Modality Self-Play for Training Multimodal LLMs Without Paired Cross-Modal Data (NeurIPS 2026 · similarity 0.88)
- ElasticMM: Efficient Multimodal LLMs Serving with Elastic Multimodal Parallelism (NeurIPS 2025 · similarity 0.88)
- Heterogeneous Parallelism for Multimodal Large Language Model Training (NeurIPS 2026 · similarity 0.87)
- Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 · 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.
