Research map / Vision-language and multimodal / Multimodal LLMs

GeoViS: Geospatially Rewarded Visual Search for Remote Sensing Visual Grounding

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 GeoViS, an MCTS-based, reward-guided visual search framework with a unified VisualRAG MLLM (fine-tuned from Qwen2.5-VL-3B) that searches for a candidate subregion and then conditionally grounds the target, reaching SOTA on five remote sensing grounding benchmarks.

Problem

Visual grounding in remote sensing imagery is hard because targets are tiny within kilometer-scale scenes and queries involve complex geospatial relations, which one-shot MLLM prediction handles poorly.

Approach

Reformulates remote sensing grounding as MCTS-based search over image subregions (zoom-in on 3x3 grids, zoom-out) guided by a geospatial reward combining QA-based semantic verification and an IoU-based centrality term, followed by conditional grounding using the global image plus the found subregion; the same MLLM supplies reward, action, and grounding, trained on data built from existing benchmar

Key results

• GeoViS reaches 79.8% Pr@0.5 on DIOR-RSVG and 68.5% Pr@0.5 on VRSBench, beating the strongest RS MLLM by over 5 points on VRSBench.
• On OPT-RSVG GeoViS achieves 70.3% Pr@0.5 versus 43.9% for Qwen3-VL-4B; on RSVG-HR it reaches 51.5% Pr@0.5.
• Trained only on DIOR-RSVG, GeoViS reaches 49.0% Pr@0.5 on VRSBench versus 39.3% for fine-tuned Qwen2.5-VL-3B.

Surprising findings

• On GeoChat, GeoViS (23.7%) is only marginally above the zero-shot-style baselines, and specialist models LQVG and LPVA are not compared on all benchmarks.
• Reward weight alpha=0.1 is optimal, meaning the IoU-based reward dominates the QA reward in the final combination.

Limitations

• Relies on MCTS with multiple MLLM calls per query, adding inference cost (10 simulations, depth 5).
• Evaluated only with a 3B backbone; comparisons to specialist models are incomplete across datasets.
• Training set expansion details (size of atomically constructed data) are not stated.

Experimental setup

Datasets: DIOR-RSVG RSVG-HR OPT-RSVG VRSBench GeoChat

Benchmarks: DIOR-RSVG VRSBench GeoChat RSVG-HR OPT-RSVG

Metrics: Pr@0.5 Pr@0.7 meanIoU

Baselines: LQVG LPVA MiniGPT-v2 Qwen2.5-VL-3B GLM-4.1V-9B-Thinking Qwen3-VL-4B EarthDial VHM GeoChat SkySenseGPT EarthGPT GeoGround

Builds on: Qwen2.5-VL LLaMA-Factory V* DyFo

Closest papers by meaning

• GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing (ICML 2025 · similarity 0.90)
• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.90)
• GeoChat: Grounded Large Vision-Language Model for Remote Sensing (CVPR 2024 · similarity 0.90)
• SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.90)
• TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation (CVPR 2026 · similarity 0.89)
• GeoDiT: A Diffusion-based Vision-Language Model for Geospatial Understanding (CVPR 2026 · similarity 0.89)