DASH: Detection and Assessment of Systematic Hallucinations of VLMs

Conference: ICCV 2025
arXiv: 2503.23573
Code: https://YanNeu.github.io/DASH
Area: Multimodal VLM
Keywords: Vision-language models, object hallucination, systematic error detection, diffusion model optimization, large-scale benchmark

TL;DR

This paper proposes DASH, a fully automated pipeline that systematically discovers false-positive object hallucination clusters in VLMs via two complementary strategies: LLM-based text query generation (DASH-LLM) and diffusion model optimization-based image query generation (DASH-OPT). Applied to ReLAION-5B, DASH uncovers 19k+ clusters and 950k+ images, and constructs the more challenging DASH-B benchmark.

Background & Motivation

VLMs (e.g., PaliGemma, LLaVA-NeXT) achieve strong performance on multimodal tasks, yet suffer from object hallucination—incorrectly identifying objects as present when they do not appear in the image. Existing benchmarks (POPE, AMBER) exhibit two fundamental limitations:

Overly restricted datasets: They rely on small-scale annotated datasets such as MSCOCO (only 80 object categories), failing to reflect VLM behavior in open-world deployment. The TNR of current models on POPE has reached 96%, approaching saturation.

Lack of systematic evaluation: They cannot distinguish between sporadic hallucinations and systematic model deficiencies. If a VLM consistently hallucinates on specific image types, this indicates a fundamental flaw requiring correction.

• Key Challenge: VLMs are deployed in open-world settings, yet evaluation benchmarks remain confined to closed datasets. Exhaustively testing all image–object combinations in ReLAION-5B is infeasible, necessitating an effective search strategy to surface systematic hallucinations.
• Core Idea: Construct a fully automated pipeline, DASH, that generates targeted queries (textual or visual) to retrieve real images from large-scale datasets that trigger VLM hallucinations, and employs clustering to identify semantically coherent systematic error patterns.

Method

Overall Architecture

DASH consists of four stages: query generation (DASH-LLM or DASH-OPT) → Exploration (kNN retrieval) → Exploitation (expanded retrieval) → Clustering. The entire pipeline is fully automated and requires no manual annotation.

Key Designs

1. DASH-LLM (Text-based Query Generation):

   • Function: Employs an LLM (Llama 3.1-70B) to generate 50 text queries per object category.
   • Mechanism: The LLM is prompted to produce scene descriptions likely to cause VLM misidentification (e.g., "fireboat" may induce hallucination of "water cannon"), without explicitly mentioning the target object.
   • Design Motivation: False-positive hallucinations commonly arise from object co-occurrence associations (e.g., Christmas decorations → Baumkuchen cake). LLMs trained on large-scale text corpora effectively capture such associations.
   • Limitation: Queries are model-agnostic (not tailored to model-specific errors), and CLIP-based retrieval from text queries may fail to surface sufficiently relevant images.

2. DASH-OPT (Optimization-based Image Query Generation):

   • Function: Optimizes the input variables of a diffusion model to synthesize images that simultaneously fool the VLM while not containing the target object.
   • Mechanism: A single-step diffusion model (distilled LDM) is used, jointly optimizing two objectives:
     • VLM loss: \(L_{\text{vlm}}(C) = -\log p_{\text{vlm}}(\text{"Yes"} \mid q(C), \text{qstnOBJ})\), maximizing the probability that the VLM responds "Yes."
     • Detector loss: \(L_{\text{det}}(C) = -\log(1 - p_{\text{det}}(\text{OBJ} \mid q(C)))\), minimizing the open-world detector (OWLv2) confidence for the target object.
     • Overall objective: \(\min_C L_{\text{vlm}}(C) + L_{\text{det}}(C)\)
   • Design Motivation: Optimizing directly in pixel space produces adversarial examples rather than natural images. Optimizing in the latent space of a diffusion model constrains generated images to lie on the natural image manifold. Unlike DASH-LLM, DASH-OPT is model-specific, enabling discovery of more unexpected hallucination patterns.

3. Exploration–Exploitation–Clustering Pipeline:

   • Exploration: For each query, kNN retrieval (CLIP similarity) is performed over ReLAION-5B to obtain candidate images; images where the detector identifies the target object or where no VLM hallucination is triggered are filtered out.
   • Exploitation: Successful images from the exploration stage undergo a second kNN retrieval (50 neighbors per image) to verify whether hallucinations transfer to semantically similar images; DreamSim is used to remove near-duplicates.
   • Clustering: Hierarchical clustering (average linkage) in the CLIP embedding space merges similar pre-clusters.

Loss & Training

• OWLv2 is applied with a very low detection threshold (conservative strategy) to ensure that images labeled as "not containing the target object" are as reliable as possible.
• Human verification shows that only 5.2% of DASH images actually contain the target object, compared to a 25.5% label noise rate in POPE.
• Fine-tuning strategy: for each object, 200 DASH images (training the model to respond "No") and 400 positive samples (training the model to respond "Yes") are sampled.

Key Experimental Results

Main Results

Model	Method	Total Images	Total Clusters	Avg. Clusters/Object	Avg. Images/Cluster
PaliGemma	DASH-LLM	99.3K	1892	5.0	52.5
PaliGemma	DASH-OPT	221.7K	3895	10.3	56.9
LLaVA-NeXT Vicuna	DASH-LLM	162.4K	3632	9.6	44.7
LLaVA-NeXT Vicuna	DASH-OPT	252.0K	4632	12.2	54.4
LLaVA-NeXT Mistral	DASH-OPT	133.8K	3229	8.5	41.5

Ablation Study / Transfer & Benchmark

Configuration	DASH-B Acc.	DASH-B TNR	POPE TNR	Notes
PaliGemma2-3B	68.9%	40.9%	97.3%	DASH-B is far harder than POPE
Ovis2-8B	71.4%	44.8%	94.9%	Hallucination remains severe
LLaVa-OneVision	75.1%	60.1%	95.8%	—
GPT-4o-mini	86.3%	76.7%	—	Strongest model still shows 23% false positives
PaliGemma (before fine-tuning)	56.4%	—	87.2%	—
PaliGemma (after fine-tuning)	68.0% (+11.6%)	—	86.4%	Fine-tuning on DASH data is effective

Key Findings

• DASH-OPT substantially outperforms DASH-LLM: it discovers approximately twice as many clusters and more diverse hallucination patterns (e.g., "leopard print" → leopard, "cartoon frog" → dam).
• Hallucinations transfer across models: images discovered via PaliGemma transfer to LLaVA-NeXT Vicuna at a rate of 43–49%.
• LLM backbone significantly affects hallucination rate: Vicuna > Mistral > Llama, positively correlated with the model's overall "Yes" tendency.
• Visual encoder has a larger impact: CLIP (60% transfer rate) > SigLIP (43%).
• Model scale has limited impact: Qwen2-VL 7B (19%) vs. 72B (18%), though the 72B model exhibits fewer hallucinations at the same TPR.
• POPE benchmark is saturated: current VLMs achieve 96% TNR on POPE, while TNR on DASH-B is only 48.6%, revealing a large volume of previously overlooked hallucinations.

Highlights & Insights

1. Open-world perspective: First systematic evaluation of VLM hallucinations on a web-scale dataset (ReLAION-5B, 5 billion images), demonstrating that existing benchmarks severely underestimate the problem.
2. Fully automated, annotation-free pipeline: A detector replaces manual verification, achieving an error rate (5.2%) far below the label noise of POPE (25.5%).
3. Elegant design of DASH-OPT: A distilled single-step diffusion model enables computationally efficient generation; latent-space optimization ensures natural rather than adversarial images.
4. Practical value: Discovered systematic hallucination patterns can be directly used for fine-tuning (DASH-B accuracy improves by 11.6%).

Limitations & Future Work

1. Although ReLAION-5B is large, it does not cover all natural image distributions; some scenarios may lack sufficient semantic neighbors to form clusters.
2. The conservative detector threshold may introduce bias when evaluating state-of-the-art VLMs whose capabilities approach those of the detector.
3. The current fine-tuning strategy is task-isolated and not integrated into the standard VLM training workflow (e.g., curriculum learning).
4. Only false-positive hallucinations (incorrectly responding "Yes") are addressed; false-negative hallucinations (incorrectly responding "No") are not systematically studied.

Related Work & Insights

• Relation to Spurious ImageNet: The latter investigates spurious correlations in image classifiers; DASH extends this concept to VLMs in open-world settings.
• Relation to DiG-IN: DiG-IN also employs optimization-guided image generation to debug models, but uses a multi-step diffusion process and targets only classifier discrepancies.
• Insight: The discovery of systematic hallucinations suggests that strong co-occurrence biases exist in VLM training data—future work may address this fundamentally through data debiasing at the pre-training stage.

Rating

• Novelty: ⭐⭐⭐⭐⭐ — A novel paradigm for systematic hallucination detection; DASH-OPT is elegantly designed.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ — 380 object categories, multiple VLMs, human verification, transfer analysis, and fine-tuning experiments.
• Writing Quality: ⭐⭐⭐⭐ — Clear structure and excellent visualizations; some experimental details require consulting the appendix.
• Value: ⭐⭐⭐⭐⭐ — Reveals that VLM hallucination is far more severe than previously recognized; DASH-B has potential to become a new standard benchmark.

Related Papers

• [NeurIPS 2025] Systematic Reward Gap Optimization for Mitigating VLM Hallucinations
• [ICCV 2025] BabyVLM: Data-Efficient Pretraining of VLMs Inspired by Infant Learning
• [ICCV 2025] HRScene: How Far Are VLMs from Effective High-Resolution Image Understanding?
• [ICCV 2025] Mitigating Object Hallucinations via Sentence-Level Early Intervention
• [CVPR 2026] Rethinking VLMs for Image Forgery Detection and Localization