Research map / Vision-language and multimodal / Multimodal LLMs

Res-Bench: Benchmarking the Robustness of Multimodal Large Language Models to Dynamic Resolution Input

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 1 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 Res-Bench, a 14,400-sample benchmark of downsampled images across 12 resolution levels, with Spearman and ACE/RCE robustness metrics, evaluating eight MLLMs and showing preprocessing and mixed-resolution fine-tuning effects.

Problem

Existing MLLM benchmarks measure semantic accuracy but not whether performance stays stable across varying input resolutions, which dynamic-resolution models are expected to handle.

Approach

Build a resolution-controlled benchmark by downsampling high-resolution, visually dependent samples to 12 levels, and quantify resolution robustness with Spearman's correlation between resolution and accuracy plus absolute and relative continuous error (ACE/RCE) for fluctuation.

Key results

• Gemini1.5-Pro achieves the highest average accuracy at 0.561, while Qwen2.5-VL (native resolution) reaches 0.547 but has the worst RCE of 0.547.
• MiniCPM-o-2.6 shows a perfectly monotonic trend (Spearman rho = 1.000) but lower average accuracy of 0.510.
• Fine-tuning Qwen2.5-VL-3B on 7,536 mixed-resolution spatial reasoning instances raised Instance Reasoning accuracy from 0.454 to 0.607 and lowered its RCE from 0.539 to 0.478.

Surprising findings

• Zero-padding low-resolution images to moderate sizes improves accuracy even though no new information is added, while excessive padding hurts.
• Native dynamic-resolution models achieve higher peak accuracy but are less robust than patch-based models.

Limitations

• Resolutions only reach about 1.5K; 4K and 8K inputs are not evaluated.
• Super-resolution is limited to a single diffusion-based method (DiffIR).
• Proprietary models were run once, and the fine-tuning study covers only one 3B model on spatial reasoning tasks.

Experimental setup

Datasets: SEED AI2D HR-Bench NaturalBench SpatialEval MMMU MMMU-Pro ChartQA-Pro BLINK MathVerse MathVista OCRBench

Benchmarks: Res-Bench

Metrics: accuracy average accuracy Spearman's rho ACE RCE

Baselines: GPT-4o Gemini-1.5 Pro Kimi-VL LLaVA-OneVision InternVL-2.5 MiniCPM-o-2.6 mPLUG-Owl3

Builds on: VLMEvalKit DiffIR

Closest papers by meaning

• Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models (AAAI 2025 · similarity 0.90)
• SEED-Bench: Benchmarking Multimodal Large Language Models (CVPR 2024 · similarity 0.89)
• II-Bench: An Image Implication Understanding Benchmark for Multimodal Large Language Models (NeurIPS 2024 · similarity 0.88)
• MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks (ICLR 2025 · similarity 0.88)
• MMBench-Live: A Continuously Evolving Benchmark for Multimodal Models (ICML 2026 · similarity 0.88)
• MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans? (ICLR 2025 · similarity 0.88)