Research map / Vision-language and multimodal / Multimodal LLMs

AgroOmni: A Large-Scale Multi-view Agricultural Dataset for Cross-Scale Multimodal Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 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 AgroOmni, a 288K-QA multi-view (Ground/UAV/Satellite) agricultural VQA training corpus, and AgroNVILA, an NVILA-Lite-8B model SFT'd on it and then GRPO-tuned, reaching 62.32% on AgroMind.

Problem

Existing agricultural MLLMs are biased toward ground-level close-up imagery and lack multi-view (UAV and satellite) VQA training data, causing scale confusion when reasoning over macro-scale agricultural scenes.

Approach

Aggregates ground, UAV and satellite agricultural images from 13 public datasets plus a proprietary parcel collection, converts labels into VQA pairs via rule-based templates and evidence-based logic synthesis, and fine-tunes NVILA-Lite-8B on it followed by GRPO with task-aware rewards.

Key results

• AgroNVILA (SFT+GRPO) reaches 62.32% overall on AgroMind, vs 47.29% for GPT-5.2 and 35.68% for the NVILA-Lite-8B base.
• On AgMMU, AgroNVILA with 30% of the data reaches 78.6% average MCQ accuracy vs 64.6% zero-shot.
• SFT-100% raises UAV-view accuracy from 19.40% to 61.10% and satellite-view from 30.97% to 59.39%.

Surprising findings

• Spatial Perception gains remain constrained (only +9.00% from 33% to 45.31%) even as other categories saturate, which the authors attribute to frozen vision encoders.
• Scaling from 33% to 100% of data yields only marginal gains over the 33% baseline on all four coarse task categories.

Limitations

• Spatial perception remains weak; frozen vision encoder limits macro-level spatial priors.
• Heavily skewed task and view distribution (ground 68%, UAV 7.9%).

Experimental setup

Datasets: AgroOmni AgroMind AgMMU CLCD EarthVQA GWHD OAM-TCD OilPalmUAV AVC PhenoBench IP102 ACFR

Benchmarks: AgroMind AgMMU

Metrics: accuracy (%) overall mean accuracy across tasks MCQ accuracy

Baselines: GPT-5.2 GPT-4o Claude-4.6-Sonnet Gemini-1.5-Pro LLaVA-NeXT-8B InternVL-2-4B GeoChat AgriGPT-VL NVILA-Lite-8B PPO DRPO

Builds on: NVILA GRPO PPO AgroMind

Closest papers by meaning

• AgMMU: A Comprehensive Agricultural Multimodal Understanding Benchmark (NeurIPS 2025 · similarity 0.94)
• Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.90)
• MIRAGE: A Benchmark for Multimodal Information-Seeking and Reasoning in Agricultural Expert-Guided Conversations (NeurIPS 2025 · similarity 0.88)
• Empowering and Assessing the Utility of Large Language Models in Crop Science (NeurIPS 2024 · similarity 0.86)
• VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing (NeurIPS 2026 · similarity 0.86)
• MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models (AAAI 2026 · similarity 0.86)