Research map / Vision-language and multimodal / Multimodal LLMs
Reason with Thumbnails, Answer with Focus: An Efficient and Effective Paradigm for Multimodal Grounded Visual Reasoning
An-Lan Wang, Guozhi Tang, Lei Liao, Hanshen Zhu, Kai Huang, Jingqun Tang, Jiaming Zhou, Kun-Yu Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an efficient grounded visual reasoning paradigm for multimodal LLMs that reasons over thumbnails and answers with focused regions, aiming to cut the cost of multi-round ground-then-answer methods (abstract only; details not available).
Problem
Grounded Visual Reasoning (GVR) improves interpretability of multimodal LLMs by selecting relevant image regions before answering, but its multi-round ground-then-answer and reasoning process is much more computationally expensive than non-GVR methods.
Approach
Reason over low-resolution thumbnails and answer with focused high-resolution regions to get efficient and effective GVR; the abstract does not give further details of the mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RTAF paradigm (thumbnail reasoning, high-res crop answering) to Ground-R1 (ablation) | Ground-R1 · VisCoT test average | average visual tokens: 847 -> 627 | Improved |
| add RTAF paradigm to Ground-R1 (ablation) | Ground-R1 · VisCoT test average | accuracy: 77.8 -> 76.2 | Hurt |
| add new reward functions to RTAF (ablation) | RTAF · VisCoT test average | accuracy: 76.2 -> 77.1 | Improved |
| add new reward functions to RTAF (ablation) | RTAF · VisCoT test average | average visual tokens: 627 -> 553 | Improved |
| add training data selection strategy (ablation) | RTAF with rewards · VisCoT test average | accuracy: 77.1 -> 78.3 | Improved |
| add bounding box reward (ablation) | RTAF · VisCoT test average | accuracy: 76.2 -> 76.7 | Improved |
| add calibrated efficiency reward (ablation) | RTAF · VisCoT test average | accuracy: 76.7 -> 77.1 | Improved |
| double the efficiency reward weight (2 x Reff) (ablation) | RTAF · VisCoT test average | accuracy: 77.1 -> 74.5 | Hurt |
| use class-wise data selection (Combination 3) instead of random data (ablation) | RTAF · VisCoT test average | accuracy: 77.1 -> 78.3 | Improved |
| use RTAF instead of Qwen2.5-VL-7B | Qwen2.5-VL-7B · VisCoT test average | accuracy: +5.8% accuracy, 471 vs. 391 tokens | Improved |
| use RTAF instead of Ground-R1 on HR-Bench-4K token usage | Pixel Reasoner and DeepEyes · HR-Bench-4K | average visual tokens: 13,118 vs 41,080 and 39,418 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- DeepEyes — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 1 of 1 shared result
- Pixel Reasoner — Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (NeurIPS 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.89)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.88)
- Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.88)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 2026 · similarity 0.88)
- Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models (NeurIPS 2025 · similarity 0.88)
- SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 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.
