Research map / Vision-language and multimodal / Multimodal LLMs
Learning to Inference Adaptively for Multimodal Large Language Models
Zhuoyan Xu, Khoi Duc Nguyen, Preeti Mukherjee, Saurabh Bagchi, Somali Chaterji, Yingyu Liang, Yin 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 |
|---|---|---|---|
| use probabilistic scheduler (sample exactly k switches) instead of deterministic hinge-penalty scheduler (ablation) | AdaLLaVA-L 7B · VQAv2 | accuracy at budget 0.65: 72.2 -> 72.7 | Improved |
| use probabilistic scheduler instead of deterministic for budget utilization (ablation) | AdaLLaVA-L 7B · VQAv2 | utilization (%) at budget 0.75: 83.2 -> 100.0 | Improved |
| use probabilistic scheduler instead of deterministic for latency constraint success at 0.95 budget (ablation) | AdaLLaVA-L 7B · VQAv2 | success rate (%): 96.1 -> 100.0 | Improved |
| apply 24 switches instead of 16 switches (last 16 layers) (ablation) | AdaLLaVA-L · VQAv2 | accuracy | Hurt |
| use 4-head instead of 8-head attention switch groups (ablation) | AdaLLaVA-H · VQAv2 | accuracy-latency tradeoff | No clear effect |
| use head/neuron-level switches (H) instead of layer-level switches (L) (ablation) | AdaLLaVA · VQAv2 | accuracy: about 75% -> about 76% | Improved |
| train the scheduler instead of random uniform sampling of execution plans (ablation) | AdaLLaVA-L on Mipha-3B · Mipha-3B | accuracy | Improved |
| reduce latency budget from 100% to 60% for AdaLLaVA-L | LLaVA-1.5 Vicuna-7B · TextVQA | accuracy: 57.0 -> 47.7 | Hurt |
| reduce latency budget from 100% to 60% for AdaLLaVA-L | LLaVA-1.5 Vicuna-7B · MMBench | accuracy: 63.7 -> 63.8 | No clear effect |
| reduce latency budget to 60% on MM-Vet and LLaVA-WILD | AdaLLaVA-L-7B · MM-Vet / LLaVA-WILD | score: MM-Vet 30.7 -> 23.9; LLaVA-WILD 64.2 -> 46.3 | Hurt |
| integrate AdaLLaVA-H with PruMerge+ token selection | PruMerge+ · VQAv2 | accuracy and FLOPs: 76.8 -> 76.0 at 100% budget; FLOPs 3.0T at 100% to 1.8T at 60% | Mixed |
| apply AdaLLaVA-L to Mipha-3B at 60% budget | Mipha-3B · MMBench | accuracy: 69.7 -> 64.6 | Hurt |
Closest papers by meaning
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.91)
- Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference (AAAI 2025 · similarity 0.91)
- Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification (ICLR 2025 · similarity 0.90)
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
- ReaGEN: Adaptive Generation of Structured Chains-of-Thought for Efficient Multimodal Reasoning (CVPR 2026 · similarity 0.90)
- Look-Back: Implicit Visual Re-focusing in MLLM Reasoning (AAAI 2026 · similarity 0.90)
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.
