Research map / Vision-language and multimodal / Multimodal LLMs

Vision Language Models are Biased

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 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 VLMBias, a semi-automated benchmark of counterfactual images across 7 subjects showing that frontier VLMs default to memorized prior knowledge and count poorly (17.05% accuracy) on modified familiar images.

Problem

VLMs may answer objective counting and identification questions on counterfactual images by relying on prior knowledge of popular subjects rather than visual evidence; prior bias evidence relied on adversarial Y/N prompts.

Approach

Use neutral prompts with minimally modified counterfactual images of well-known subjects (animals, logos, flags, chess, game boards, illusions, novel patterned grids), generated with image editors and code, and measure accuracy and a predefined "bias rate" that isolates prior-driven answers.

Key results

• Five SOTA VLMs average 17.05% accuracy on counterfactual counting questions while achieving 100% on unmodified images.
• Removing image backgrounds raises mean accuracy by 21.09 points (17.05% to 38.14%) and cuts bias rate by 40.58 points.
• Counting accuracy rises with reasoning tokens to about 40% before declining with overthinking; VLMs explicitly trained to count (Molmo-72B, Moondream-2B) average 36.02%.

Surprising findings

• Thinking models perform only marginally better than non-thinking models, and GPT-4o-mini with tools improves only slightly (23.18% to 25.08%) because it rarely uses tools.
• Larger VLMs (Pixtral, Qwen2.5-VL) show inverse scaling, with about 1.26x higher bias rates than smaller ones.

Limitations

• Image generators used for creating the benchmark carry their own biases, so some counterfactual edits fail.
• Human-in-the-loop filtering of generated images is required, so the framework is only semi-automated.

Experimental setup

Datasets: VLMBias

Benchmarks: VLMBias

Metrics: accuracy bias rate pass@5 agreement-based consistency

Baselines: Gemini 2.5 Flash Claude 3.5 Sonnet GPT-4.1 GPT-4o GPT-4o-mini Molmo-72B Moondream-2B Grok GPT-5

Builds on: HallusionBench IllusionVQA BlindTest LangSAM

Closest papers by meaning

• Unveiling the Visual Counting Bottleneck in Vision-Language Models (ICML 2026 · similarity 0.90)
• Understanding Counting Mechanisms in Large Language and Vision-Language Models (CVPR 2026 · similarity 0.90)
• See It from My Perspective: How Language Affects Cultural Bias in Image Understanding (ICLR 2025 · similarity 0.89)
• Caption This, Reason That: VLMs Caught in the Middle (NeurIPS 2025 · similarity 0.89)
• VHELM: A Holistic Evaluation of Vision Language Models (NeurIPS 2024 · similarity 0.89)
• Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.88)