Research map / Vision-language and multimodal / Multimodal LLMs
VisReflect: Latent Visual Reflection for Fine-Grained Perception in Long Visual Context
Xiaoqian Shen, Mohamed Elhoseiny
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 visual reflection with cosine loss (ablation) | Qwen2.5-VL Finetune · V* | accuracy: 80.1 -> 84.9 | Improved |
| add visual reflection with cosine loss (ablation) | Qwen2.5-VL Finetune · HRBench-4K | accuracy: 71.3 -> 73.8 | Improved |
| add visual reflection with cosine loss (ablation) | Qwen2.5-VL Finetune · BLINK | accuracy: 55.2 -> 57.1 | Improved |
| fine-tune backbone without visual reflection (ablation) | Qwen2.5-VL · BLINK | accuracy: 56.5 -> 55.2 | Hurt |
| use cosine similarity instead of MSE as alignment loss (ablation) | VisReflect · V* | accuracy: 83.2 -> 84.9 | Improved |
| use cosine similarity instead of MSE as alignment loss (ablation) | VisReflect · HRBench-8K | accuracy: 69.3 -> 70.1 | Improved |
| add visual supervision on reflection tokens instead of Coconut (ablation) | VisReflect · V* | accuracy: 79.4 -> 84.9 | Improved |
| increase training <VR> tokens from 16 to 128 (ablation) | VisReflect · V* | accuracy: 81.6 -> 84.9 | Improved |
| increase training <VR> tokens from 128 to 196 (ablation) | VisReflect · V* | accuracy: 84.9 -> 83.2 | Hurt |
| use one <VR> token at inference (ablation) | VisReflect · V* | accuracy: 78.53 -> 84.85 | Improved |
| increase inference <VR> tokens on images from 1 to 32 (ablation) | VisReflect · V* | accuracy: 84.85 -> 84.72 | No clear effect |
| increase inference <VR> tokens on video from 0 to 16 (ablation) | VisReflect · MVBench | accuracy: 68.44 -> 70.54 | Improved |
Closest papers by meaning
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.89)
- Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.89)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.89)
- Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception (ICML 2026 · similarity 0.89)
- VisMem: Latent Vision Memory Unlocks Potential of Vision-Language Models (CVPR 2026 · similarity 0.89)
- BVS: Bayesian Visual Search with Multimodal Large Language Model for Fine-grained Perception (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.
