Research map / Vision-language and multimodal / Multimodal LLMs

UHR-BAT: Budget-Aware Token Compression Vision-Language model for Ultra-High-Resolution Remote Sensing

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 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 UHR-BAT, a query-guided, region-wise preserve-and-merge token compression method with multi-scale budgeted tokens for ultra-high-resolution remote sensing MLLMs, evaluated on XLRS-Bench, RSHR-Bench and MME-RealWorld-RS.

Problem

Ultra-high-resolution remote sensing images produce far too many visual tokens for MLLMs, and existing downsampling, tiling, or global top-K pruning either lose small query-critical details or cost unpredictable compute.

Approach

Scores visual tokens by text-to-vision attention from an anchor-scale view, transfers the scores across scales, partitions tokens into regions, keeps above-mean tokens per region, and merges the rest into one mean-pooled token per region under a hard per-scale budget.

Key results

• On XLRS-Bench, UHR-BAT reaches 44.0 w.Avg., above LLaVA-OneVision-72B (41.1) and InternVL3-78B (43.5), and above GeoLLaVA-8K (43.3).
• On MME-RealWorld-RS, UHR-BAT achieves 33.33 average accuracy, versus 28.41 for GeoLLaVA-8K.
• On RSHR-Bench, UHR-BAT scores 29.2 Perception and 45.0 Reasoning averages, with the best Avg on remote-sensing and open-source baselines.

Surprising findings

• Downsampling 8K inputs to 4K gives performance on par with or slightly above 8K under the same token budget (about 0.1%).
• Accuracy saturates quickly as the token budget grows, with little gain beyond a few thousand tokens.

Limitations

• Gains over GeoLLaVA-8K on XLRS-Bench are small (+0.7%), and performance is mixed across sub-tasks.
• Region partitioning via SAM incurs prohibitive preprocessing latency, so the authors default to clustering.
• Budgets are set per benchmark and resolution, and the 8K-to-4K comparison is limited to a single model.

Experimental setup

Datasets: SuperRS-VQA HighRS-VQA

Benchmarks: XLRS-Bench RSHR-Bench MME-RealWorld-RS

Metrics: accuracy weighted average accuracy

Baselines: GeoLLaVA-8K GeoChat EarthDial VHM Qwen2.5-VL-7B InternVL3-78B LLaVA-OneVision-72B GPT-4o Claude 3.7 Sonnet VILA-HD

Builds on: LLaVA-NeXT LongVA CLIP ViT-L/14 SAM K-means

Closest papers by meaning

• ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks (CVPR 2026 · similarity 0.88)
• GeoLLaVA-8K: Scaling Remote-Sensing Multimodal Large Language Models to 8K Resolution (NeurIPS 2025 · similarity 0.88)
• Global Compression Commander: Plug-and-Play Inference Acceleration for High-Resolution Large Vision-Language Models (AAAI 2026 · similarity 0.86)
• Prune Redundancy, Preserve Essence: Vision Token Compression in VLMs via Synergistic Importance-Diversity (ICLR 2026 · similarity 0.86)
• Remote Sensing Vision-Language Foundation Models without Annotations via Ground Remote Alignment (ICLR 2024 · similarity 0.85)
• PVC: Progressive Visual Token Compression for Unified Image and Video Processing in Large Vision-Language Models (CVPR 2025 · similarity 0.85)