Research map / Vision-language and multimodal / Multimodal LLMs
Diffusion-Based Immersive Visual Reasoning
Yifeng Zhang, Ming Jiang, Qi Zhao
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 LHR latent hypothesis refiner (with RGO) (ablation) | DIVR · ZInD Hard | accuracy (%): 39.44 -> 41.89 | Improved |
| add LHR latent hypothesis refiner (with RGO) (ablation) | DIVR · IQVA | accuracy (%): 54.24 -> 56.75 | Improved |
| add learned RGO instead of saliency viewports (ablation) | DIVR with LHR · IQVA | accuracy (%): 53.46 -> 56.75 | Improved |
| add learned RGO instead of CLIP-similarity selection (ablation) | DIVR with LHR · ZInD Hard | accuracy (%): 38.68 -> 41.89 | Improved |
| use dynamic CLIP viewport selection instead of static saliency (ablation) | DIVR with LHR · IQVA | accuracy (%): 53.46 -> 55.14 | Improved |
| add LHR when using saliency viewports (ablation) | DIVR · 360+X | accuracy (%): 56.72 -> 57.71 | Hurt |
| increase diffusion steps from 5 to 20 (ablation) | DIVR · ZInD Hard | accuracy (%): 38.46 -> 41.89 | Improved |
| increase diffusion steps from 5 to 20 (ablation) | DIVR · ZInD Hard | speed (QPS): 3.57 -> 1.08 | Hurt |
| use DIVR active loop instead of saliency input | Qwen2.5-VL-7B · IQVA | accuracy (%): 53.22 -> 56.75 | Improved |
| use cubemap instead of equirectangular input | Qwen2.5-VL-7B · IQVA | accuracy (%): 43.54 -> 49.20 | Improved |
| use DIVR viewport selection instead of saliency viewports | DIVR · IQVA fixation | NSS: 0.36 -> 0.49 | Improved |
| DIVR vs GPT-5 with saliency viewports | DIVR + Qwen2.5-VL-7B · ZInD Hard | accuracy (%): 35.64 -> 41.89 | Improved |
Closest papers by meaning
- PhysVLM-AVR: Active Visual Reasoning for Multimodal Large Language Models in Physical Environments (NeurIPS 2025 · similarity 0.89)
- DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.89)
- UniVR: Thinking in Visual Space for Unified Visual Reasoning (NeurIPS 2026 · similarity 0.88)
- Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · similarity 0.88)
- The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.88)
- VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 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.
