Research map / Vision-language and multimodal / Multimodal LLMs

Unveiling the Tapestry of Consistency in Large Vision-Language Models

NeurIPS 2024 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 2 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 ConBench, a benchmark of 1K images with true/false, multiple-choice, and VQA prompts around the same knowledge point plus captions, measuring answer consistency across solution-space sizes in 14 LVLMs, and proposes training-free trigger-based diagnostic refinement to improve caption consistency.

Problem

LVLMs often give inconsistent answers about the same knowledge point when the prompt's solution space differs (true/false vs. multiple-choice vs. open VQA), and existing benchmarks use single prompt formats, so they do not measure this consistency.

Approach

Build a benchmark where each image has three discriminative question types (T/F, MC, limited VQA) around one knowledge point plus a caption prompt, score consistency across them (ConScore[D]) and between captions and discriminative answers judged by GPT-4 (ConScore[C]), and use low-confidence words to trigger self-verification questions that refine captions without training.

Key results

• Qwen-VL-Max tops ConScore[D] at 37.0 accuracy-based leaderboard; GPT-4-Omni leads ConScore[C] with 55.0 and is the only model above 60 in the consistency-with-caption metric.
• Trigger-based diagnostic refinement raises LLaVA-NeXT-34B ConScore[C] by 9.1 points (to 57.4) and MiniGemini-34B by 9.6 points (to 60.2).
• Pearson coefficients between discriminative accuracy and caption consistency are all above 0.85 across figures.

Measured effects reported in the paper

Surprising findings

• Cases where all three question types are answered wrong yet remain consistent are very rare (below 0.50%), suggesting failures are inconsistent rather than systematically shared.
• Closed-source models lie above the consistency-accuracy regression line, showing a bias advantage in consistency beyond accuracy.

Limitations

• Caption-discriminative consistency is judged by GPT-4 with about 95% manual agreement, risking inaccurate evaluation.
• Trigger-based refinement adds computational cost and is bounded by the base LVLM's capabilities.
• Benchmark covers only 1K images, with some question types generated by GPT and reviewed manually.

Experimental setup

Datasets: MME SEED-Bench MMBench MMMU

Benchmarks: ConBench

Metrics: ConScore[D] ConScore[C] accuracy Pearson correlation

Baselines: GPT-4V GPT-4-Omni Gemini-Pro-Vision Gemini-Ultra-Vision Qwen-VL-Plus Qwen-VL-Max LLaVA-v1.5-7B LLaVA-NeXT-34B MiniGemini-34B InternVL-v1.2P-40B

Builds on: MME MMBench SEED-Bench MMMU ANLS

Closest papers by meaning

• Consistency and Uncertainty: Identifying Unreliable Responses From Black-Box Vision-Language Models for Selective Visual Question Answering (CVPR 2024 · similarity 0.91)
• SC-Tune: Unleashing Self-Consistent Referential Comprehension in Large Vision Language Models (CVPR 2024 · similarity 0.91)
• Natural Language Inference Improves Compositionality in Vision-Language Models (ICLR 2025 · similarity 0.91)
• Prism: A Framework for Decoupling and Assessing the Capabilities of VLMs (NeurIPS 2024 · similarity 0.90)
• VP-Bench: A Comprehensive Benchmark for Visual Prompting in Multimodal Large Language Models (AAAI 2026 · similarity 0.90)
• Seeing but Not Believing: Probing the Disconnect Between Visual Attention and Answer Correctness in VLMs (ICLR 2026 · similarity 0.90)