Research map / Vision-language and multimodal / Multimodal LLMs
Foveated Reasoning: Stateful, Action-based Visual Focusing for Vision-Language Models
Juhong Min, Lazar Valkov, Vitali Petsiuk, Hossein Souri, Deen Dayal Mohan
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 coldstart foveation SFT (ablation) | Qwen2.5-VL-7B 336^2 · DocVQA | Accuracy: 55.1 -> 64.2 | Improved |
| add coldstart foveation SFT (ablation) | Qwen2.5-VL-7B 336^2 · TextVQA | Accuracy: 77.2 -> 82.2 | Improved |
| add coldstart foveation SFT (ablation) | Qwen2.5-VL-7B 336^2 · Visual7W | Accuracy: 46.5 -> 58.3 | Improved |
| add RL (lambda_reg 1.0) after coldstart (ablation) | FoveateR-7B coldstart · DocVQA | Accuracy: 64.2 -> 61.5 | Hurt |
| add RL (lambda_reg 1.0) after coldstart (ablation) | FoveateR-7B coldstart · TextVQA | Accuracy: 82.2 -> 81.2 | Mixed |
| add RL (lambda_reg 1.0) after coldstart (ablation) | FoveateR-7B coldstart · Visual7W | Accuracy: 58.3 -> 57.2 | Mixed |
| use action-based single-pass foveation instead of multi-pass VisCoT | FoveateR-7B · DocVQA (336^2) | Accuracy: 47.6 -> 83.3 | Improved |
| use action-based single-pass foveation instead of multi-pass VisCoT | FoveateR-7B · Visual CoT (336^2) | Avg visual tokens: 1,152 -> 322.5 | Improved |
| use FoveateR instead of VisCoT | FoveateR-7B · Flickr30k (336^2) | Accuracy: 66.8 -> 60.8 | Hurt |
| increase low-res input from 224^2 to 336^2 | FoveateR-7B · DocVQA | Accuracy: 73.9 -> 83.3 | Improved |
| scale backbone from 3B to 7B | FoveateR · DocVQA (336^2) | Accuracy: 74.0 -> 83.3 | Improved |
| use FoveateR-3B instead of VisCoT-7B | FoveateR-3B · V* Bench GPT4V-hard | Accuracy: 11.8 -> 76.5 | Improved |
Closest papers by meaning
- CoFFT: Chain of Foresight-Focus Thought for Visual Language Models (NeurIPS 2025 · similarity 0.90)
- BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception (NeurIPS 2025 · similarity 0.90)
- The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.90)
- DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.90)
- ERGO: Efficient High-Resolution Visual Understanding for Vision-Language Models (ICLR 2026 · similarity 0.90)
- DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 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.
