Research map / Vision-language and multimodal / Multimodal LLMs
Scaling Inference-Time Search with Vision Value Model for Improved Visual Comprehension
Xiyao Wang, Zhengyuan Yang, Linjie Li, Hongjin Lu, Yuancheng Xu, Chung-Ching Lin, Kevin Lin, Furong Huang, Lijuan Wang
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 |
|---|---|---|---|
| guide search with VisVM instead of CLIP score | LLaVA-Next-7B search · COCO Val2014 CHAIRs | CHAIRs: 28.4 -> 26.2 | Improved |
| guide search with VisVM instead of greedy decoding | LLaVA-Next-7B · COCO Val2014 CHAIRs | CHAIRs: 32.4 -> 26.2 | Improved |
| guide search with VisVM instead of greedy decoding | LLaVA-Next-7B · MMHal | MMHal score: 2.94 -> 3.30 | Improved |
| use VisVM instead of MCTS search | LLaVA-Next-7B · COCO Val2014 CHAIRs | CHAIRs: 25.9 vs 26.2 | No clear effect |
| train VisVM with SigLIP PRM instead of CLIP PRM (ablation) | VisVM · COCO Val2014 CHAIRs | CHAIRs: 26.2 -> 25.6 | Improved |
| select candidates with VisVM instead of CLIP from same candidates (ablation) | LLaVA-Next-7B · COCO Val2014 CHAIRs | CHAIRs: 31.6 -> 30.9 | Improved |
| increase search step size for VisVM-guided search (ablation) | VisVM · COCO Val2014 CHAIRs | CHAIRs | Improved |
| apply VisVM to Qwen2-VL-7B | Qwen2-VL-7B · COCO Val2014 CHAIRs | CHAIRs: 30.8 -> 24.5 | Improved |
| self-train on VisVM-search captions instead of greedy captions | LLaVA-Next-Mistral-7B · multimodal benchmark average | average improvement: -1.6% -> +10.8% | Improved |
| self-train on greedy-decoding captions | LLaVA-Next-Mistral-7B · multimodal benchmark average | average improvement: -1.6% | Hurt |
| self-train on VisVM captions instead of CLIP-PRM captions | Qwen2-VL-7B-Instruct · multimodal benchmark average | average improvement: +5.0% -> +7.3% | Improved |
| self-train on VisVM captions | LLaVA-Next-Mistral-7B · COCO Val2014 CHAIRs | CHAIRs: 32.4 -> 22.6 | Improved |
Closest papers by meaning
- Dual-Stage Value-Guided Inference with Margin-Based Reward Adjustment for Fast and Faithful VLM Captioning (NeurIPS 2025 · similarity 0.91)
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.89)
- SearchLVLMs: A Plug-and-Play Framework for Augmenting Large Vision-Language Models by Searching Up-to-Date Internet Knowledge (NeurIPS 2024 · similarity 0.89)
- Enhancing Large Vision Language Models with Self-Training on Image Comprehension (NeurIPS 2024 · similarity 0.89)
- VisMem: Latent Vision Memory Unlocks Potential of Vision-Language Models (CVPR 2026 · similarity 0.89)
- TGV-KV: Text-Grounded KV Eviction for Vision-Language Models (ICML 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.
