Research map / Vision-language and multimodal / Multimodal LLMs
Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning
Yinan ZHOU, Haokun Lin, Yichen Wu, Yuxin Chen, Teng Wang, Caifeng Shan, Zhenan Sun, Chen Ma, Li Zhu, Ying Shan
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 JRL routing instead of random routing at 60% to draft | Qwen2.5-VL-7B + Skywork-R1V-38B · MathVista | accuracy: 71.6% -> 74.2% | Improved |
| use JRL routing instead of token-probability routing | Qwen2.5-VL-7B + Skywork-R1V-38B · MathVerse (50% to draft) | accuracy: 56.9% -> 58.5% | Improved |
| use JRL routing instead of special-token routing | Qwen2.5-VL-7B + Skywork-R1V-38B · MathVista (60% to draft) | accuracy: 72.1% -> 74.2% | Improved |
| use DRL (draft rating only) routing | Qwen2.5-VL-7B draft · MathVerse | routing AUC (%): 56.13 -> 55.44 | Hurt |
| add target rating (JRL) on top of DRL (ablation) | Draft model · MathVerse | routing AUC (%): 55.44 -> 56.97 | Improved |
| add target rating (JRL) on top of DRL (ablation) | Draft model · ChartQA | routing AUC (%): 81.97 -> 83.61 | Improved |
| add target rating (JRL) on top of DRL (ablation) | Draft model · MathVista | routing AUC (%): 73.29 -> 73.36 | No clear effect |
| add Internal Probability Ranking within score bands (ablation) | DRL/JRL routing · ChartQA and MathVista | routing quality | Improved |
| replace SFT-based rating training with RL-based rating training (ablation) | Draft model · MathVista | draft accuracy: drops by 4.2% with SFT | Improved |
| raise rating loss weight from 1.5 to 5 or 10 (ablation) | DRL/JRL training · MathVerse | accuracy | Hurt |
| use larger tau for JRL score (ablation) | JRL routing · MathVista | routing curve | Mixed |
| use DRL-trained draft for routing on ChartQA with target 84.39 | DRL · ChartQA (p_t>40) | accuracy: 84.39 -> 84.72 | Improved |
Closest papers by meaning
- Dual-Latent Memory Routing for Vision-Language Reasoning (ICML 2026 · similarity 0.90)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.90)
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.89)
- DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.89)
- See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs (CVPR 2026 · similarity 0.89)
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.
