When Prompts Override Vision: Prompt-Induced Hallucinations in LVLMs

2026-04-23 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Natural Language Processing · Depth: Expert, quick

Summary

A new benchmark called HalluScope has been introduced to investigate prompt-induced hallucinations in large vision-language models (LVLMs). This research identifies that LVLM hallucinations primarily arise from an over-reliance on textual priors and background knowledge, particularly when such information is embedded within textual instructions. To address this, the authors propose HalluVL-DPO, a fine-tuning framework that uses preference optimization. HalluVL-DPO guides LVLMs to generate more visually grounded responses by leveraging a specially curated training dataset. The optimized model effectively reduces targeted hallucination failures while maintaining or enhancing performance on existing hallucination benchmarks and visual capability assessments. The benchmark, training dataset, and code will be publicly released to foster further research.

Key takeaway

For AI Engineers and Research Scientists developing or deploying LVLMs, understanding that textual prompt priors significantly induce hallucinations is critical. You should consider integrating HalluVL-DPO or similar preference optimization techniques into your fine-tuning workflows to mitigate these prompt-induced hallucinations, ensuring your models produce more visually accurate and reliable outputs. Evaluate your models using benchmarks like HalluScope to specifically identify and address these failure modes.

Key insights

LVLM hallucinations are largely driven by over-reliance on textual instruction priors, not just vision backbone limits.

Principles

• Textual priors can override visual input.
• Preference optimization improves visual grounding.

Method

HalluVL-DPO fine-tunes LVLMs using preference optimization on a curated dataset, guiding models to prefer visually grounded responses over hallucinated ones.

In practice

• Use HalluScope to evaluate prompt-induced hallucinations.
• Apply HalluVL-DPO for visually grounded LVLM responses.

Topics

• Large Vision-Language Models
• Hallucinations
• HalluScope Benchmark
• Textual Priors
• Preference Optimization

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Mechanisms of Prompt-Induced Hallucination in Vision-Language Models — VLMs hallucinate by prioritizing text over vision, a behavior traceable to specific, ablatable attention heads.
• Reducing Hallucination in Vision-Language Models via Stage-wise Preference Optimization under Distribution Shift — Stage-wise preference optimization with targeted data construction effectively reduces VLM hallucination by focusing on difficult grounding boundaries.
• Your LLM Has Hallucination Neurons. There Are Only a Handful of Them. A MUST READ — A recent finding identifies "H-Neurons," a minuscule subset comprising less than 0.1% of total neurons, as the causal drivers of hallucinations in large language models.
• LLM-as-Judge Framework for Evaluating Tone-Induced Hallucination in Vision-Language Models — Linguistic tone systematically influences VLM hallucination, with varying effects on frequency and intensity across models and tasks.
• [Demo] I found a way to physically break LLM hallucinations using "Visual Anchors" (Modality Shift) — Inserting visual data into an LLM's context can prevent hallucinations by forcing a modality shift.
• The Mirage of Performance Gains: Why Contrastive Decoding Fails to Mitigate Object Hallucinations in MLLMs? — Contrastive decoding's MLLM hallucination mitigation gains are spurious, driven by output bias and sampling degradation, not true reduction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.