Research map / Vision-language and multimodal / Multimodal LLMs

GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 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.

Introduces GeoMMBench, a 1,053-question expert-crafted multimodal benchmark for geoscience and remote sensing, evaluates 36 MLLMs and LLMs, and proposes GeoMMAgent, a training-free multi-agent tool-augmented framework that reaches 88.4% on the test set.

Problem

Existing remote sensing benchmarks focus on perception tasks with optical imagery and lack multidisciplinary, multi-sensor, expert-level knowledge and reasoning evaluation for MLLMs in geoscience.

Approach

An expert-authored multiple-choice benchmark spanning disciplines, sensor modalities, and a four-level task spectrum, paired with a training-free plan-execute-evaluate multi-agent system that orchestrates retrieval, RS perception tools, and a reasoning MLLM.

Key results

• GeoMMAgent reaches 88.4% accuracy on the GeoMMBench test set, versus 70.7% for the best standalone model Gemini-1.5-pro.
• Human experts score 86.5% on the val set; GeoMMAgent matches this at 86.5% on val.
• Best open-source model Qwen3-VL-30B reaches 66.7% on test; removing the reasoning module drops GeoMMAgent to 67.3% on test.

Surprising findings

• Frontier model GPT-5 scores only 46.1% on test, below several open-source models.
• Models can answer text-only questions correctly but fail when the same knowledge must be grounded in a diagram, indicating visual-encoder misalignment.

Limitations

• Benchmark is manually curated with 1,016 test questions, which may introduce selection bias and cannot cover all geoscience knowledge.
• GeoMMAgent's tools are targeted to GeoMMBench tasks and do not cover the full range of geoscience and RS tasks.
• Only 37 validation questions are used for human expert evaluation, making the human baseline small.

Experimental setup

Datasets: Million-AID DOTA-v2 LoveDA GeoMMBench

Benchmarks: GeoMMBench

Metrics: micro-averaged accuracy

Baselines: GPT-4o GPT-4v Gemini-1.5-pro Gemini-2.0-flash Qwen2.5-VL-7B Qwen3-VL-30B GeoChat VHM LHRS-Bot TeoChat

Builds on: LMMs-Eval Model Context Protocol

Closest papers by meaning

• Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.89)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.89)
• GeoCoT: Towards Reliable Remote Sensing Reasoning with Manifold Perspective (CVPR 2026 · similarity 0.89)
• SpatialScore: Towards Comprehensive Evaluation for Spatial Intelligence (CVPR 2026 · similarity 0.88)
• PEACE: Empowering Geologic Map Holistic Understanding with MLLMs (CVPR 2025 · similarity 0.88)
• SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.88)