Research map / Vision-language and multimodal / Multimodal LLMs

Are We on the Right Way for Evaluating Large Vision-Language Models?

NeurIPS 2024 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 75 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.

Identifies visual-unnecessary samples and multi-modal data leakage in existing LVLM benchmarks, then curates MMStar (1,500 vision-indispensable samples across 6 capabilities/18 axes) and proposes multi-modal gain (MG) and multi-modal leakage (ML) metrics.

Problem

Existing LVLM multi-modal benchmarks contain many samples answerable without images (via text or LLM world knowledge), and LLMs/LVLMs may have memorized evaluation samples during training, so scores misjudge true multi-modal capability and gains.

Approach

Filter benchmark samples with LLM "inspectors" (drop samples answerable by multiple text-only LLMs), then have LVLMs gauge difficulty and humans curate a balanced vision-indispensable benchmark; measure visual-free performance of LVLMs and their LLM bases to derive MG and ML metrics that quantify real multimodal gain and leakage.

Key results

• GPT4V (high-resolution) achieves the best MMStar average of 57.1%, while the best open-source model InternLM-XComposer2 reaches 55.4%.
• LLM baselines score near random on MMStar (best 20.6% for GeminiPro, 26.4% for Qwen1.5-1.8B with 2-shot), and GeminiPro reaches 42.9% on MMMU with no image input.
• Sphinx-X-MoE scores 43.6% on MMMU without images, exceeding its LLM backbone by 17.9 points; the average ML across models is 1.9 on MMStar versus higher values on other benchmarks.

Measured effects reported in the paper

Surprising findings

• GeminiPro achieves 42.9% on MMMU without any visual input and beats random choice by over 24% on average across six benchmarks.
• Sphinx-X-MoE scores 43.6% on MMMU without images, surpassing its own LLM backbone by 17.9 points.

Limitations

• Filtering relies on heuristic answer-matching rules rather than GPT-4 extraction, which may introduce scoring errors.
• The benchmark is static (1,500 samples), so it may itself leak into future training corpora, as the authors note.
• MG and ML depend on the availability of an LLM base for each LVLM, limiting applicability to models without a matched text-only backbone.

Experimental setup

Datasets: MMMU MMBench ScienceQA AI2D SEED-Bench MathVista

Benchmarks: MMStar MMMU MMBench ScienceQA AI2D SEED MathVista

Metrics: accuracy multi-modal gain (MG) multi-modal leakage (ML)

Baselines: Random Choice GPT4V GeminiPro-Vision LLaVA-1.5 InternLM-XComposer2 LLaVA-Next InternVL-Chat-v1.2 Yi-VL

Builds on: VLMEvalKit

Closest papers by meaning

• Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation (ICLR 2026 · similarity 0.93)
• MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.92)
• MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs (ICML 2026 · similarity 0.92)
• VHELM: A Holistic Evaluation of Vision Language Models (NeurIPS 2024 · similarity 0.92)
• Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.92)
• VLKEB: A Large Vision-Language Model Knowledge Editing Benchmark (NeurIPS 2024 · similarity 0.92)