FaithSCAN: Model-Driven Single-Pass Hallucination Detection for Faithful Visual Question Answering

Chaodong Tong, Qi Zhang, Chen Li, Lei Jiang, Yanbing Liu

TL;DR

FaithSCAN is a lightweight, "single-pass" hallucination detector for Visual Question Answering (VQA) that identifies ungrounded answers by analyzing a model's internal states. By fusing internal visual, textual, and cross-modal alignment signals, it achieves state-of-the-art detection accuracy while being orders of magnitude faster than traditional sampling-based methods.

What Problem Does This Paper Address?

Vision-Language Models (VLMs) are increasingly capable, but they suffer from a persistent reliability issue: faithfulness hallucinations. This occurs when a model generates a fluent, confident answer that simply isn't supported by the image (e.g., claiming there is a "blue tie" when the person isn't wearing one).

Current detection methods generally fall into two camps, both with significant drawbacks:

  1. External Verification: Using auxiliary models (like object detectors or LLM "judges") to double-check the answer. This is computationally heavy and prone to "cascading errors" if the checker model is also wrong.
  2. Uncertainty-Driven Methods: Sampling the model multiple times (e.g., generating 10 different answers) to see if it's "confused." This is incredibly slow and only looks at the final output, ignoring the rich diagnostic information hidden inside the model's layers.

The authors of FaithSCAN argue that hallucinations leave "fingerprints" within the model's internal reasoning process. They set out to build a detector that can spot these fingerprints in a single forward pass.

Methods

FaithSCAN is a lightweight supervised network designed to sit on top of a frozen VLM. Its core innovation lies in how it extracts and fuses three types of Internal Uncertainty Signals:

  • Token-level Generation Signals: It looks at the log-likelihood, entropy, and hidden states of the generated text.
  • Visual Semantic Representations: It extracts intermediate patch embeddings from the vision encoder to see how the model "perceives" the image.
  • Cross-modal Aligned Representations: It examines the features where visual and textual information are projected into a shared space.

The Architecture

The system uses Branch-wise Evidence Encoding to process these heterogeneous signals separately, followed by an Uncertainty-Aware Attention mechanism. This attention layer learns which signal to trust for a given query—for example, prioritizing visual signals for "is there a cat?" questions and textual signals for complex reasoning tasks.

Low-Cost Supervision

To train FaithSCAN without expensive human labeling, the authors used a "Model-as-a-Judge" approach. They employed a strong VLM (Qwen2.5-VL) to act as a Visual-NLI (Natural Language Inference) judge, comparing the model's generated answer against a ground-truth reference while looking at the image. This allowed them to automatically generate high-quality "hallucinated vs. faithful" labels for training.

Datasets

The authors evaluated FaithSCAN across four major VQA benchmarks, each covering different hallucination types:

  • HalLoc-VQA: Focuses on object and scene-level hallucinations with token-level labels.
  • POPE: Specifically targets object existence (e.g., "Is there a chair?").
  • HaloQuest: Emphasizes reasoning failures and "false premise" questions.
  • VQA v2: A standard, naturalistic VQA dataset.

Key Insights and Ideas

  • Hallucinations are "Structured" Phenomena: The paper demonstrates that hallucinations aren't just random noise; they arise from systematic variations in internal states across perception, reasoning, and decoding.
  • The "Sweet Spot" Layer: Through layer-wise analysis, the researchers found that intermediate layers (roughly 2/3 deep into the model) provide the best signals for detection. Early layers are too raw, and final layers are too focused on language fluency to reveal the underlying "confusion."
  • Signal Complementarity: Different signals catch different lies. Visual patch uncertainty is highly effective for detecting object hallucinations (perceptual failures), while token-level uncertainty is better for reasoning-level hallucinations.
  • Architecture Matters: Hallucination patterns vary significantly between VLM architectures (e.g., LLaVA vs. InstructBLIP), suggesting that a "one-size-fits-all" external detector is less effective than a model-driven internal one.

Key Results

  • Superior Accuracy: FaithSCAN consistently outperformed baselines like Semantic Entropy (SE) and SelfCheckGPT, achieving up to an 8% improvement in AUROC on certain benchmarks.
  • Massive Speedup: Because it only requires a single forward pass, FaithSCAN is significantly more efficient. For 100 samples, FaithSCAN's inference takes roughly 0.6 to 1.2 seconds, whereas sampling-based methods like SE can take over 1,500 seconds.
  • Interpretability: Using a "Gradient x Input" attribution method, the authors showed that FaithSCAN can pinpoint exactly which tokens in a sentence are hallucinated (e.g., highlighting the word "stool" when the model should have said "chair").

Limitations and Future Work

  • The OOD Gap: While FaithSCAN generalizes better than simple logistic regression, there is still a performance gap when moving from "in-distribution" training data to "out-of-distribution" datasets.
  • White-box Requirement: This method requires access to the model's internal weights and activations. It cannot be used for "black-box" APIs (like the standard ChatGPT/GPT-4o API) unless the provider exposes internal logit and embedding data.
  • Visual Mapping: While the model can identify which words are hallucinations, mapping those signals back to specific pixels in the input image remains a challenge for future research.

Summary updated

Abstract

Faithfulness hallucinations in VQA occur when vision-language models produce fluent yet visually ungrounded answers, severely undermining their reliability in safety-critical applications. Existing detection methods mainly fall into two categories: external verification approaches relying on auxiliary models or knowledge bases, and uncertainty-driven approaches using repeated sampling or uncertainty estimates. The former suffer from high computational overhead and are limited by external resource quality, while the latter capture only limited facets of model uncertainty and fail to sufficiently explore the rich internal signals associated with the diverse failure modes. Both paradigms thus have inherent limitations in efficiency, robustness, and detection performance. To address these challenges, we propose FaithSCAN: a lightweight network that detects hallucinations by exploiting rich internal signals of VLMs, including token-level decoding uncertainty, intermediate visual representations, and cross-modal alignment features. These signals are fused via branch-wise evidence encoding and uncertainty-aware attention. We also extend the LLM-as-a-Judge paradigm to VQA hallucination and propose a low-cost strategy to automatically generate model-dependent supervision signals, enabling supervised training without costly human labels while maintaining high detection accuracy. Experiments on multiple VQA benchmarks show that FaithSCAN significantly outperforms existing methods in both effectiveness and efficiency. In-depth analysis shows hallucinations arise from systematic internal state variations in visual perception, cross-modal reasoning, and language decoding. Different internal signals provide complementary diagnostic cues, and hallucination patterns vary across VLM architectures, offering new insights into the underlying causes of multimodal hallucinations.
BibTeX
@misc{tong2026faithscan,
  author = {Chaodong Tong and Qi Zhang and Chen Li and Lei Jiang and Yanbing Liu},
  title = {{FaithSCAN: Model-Driven Single-Pass Hallucination Detection for Faithful Visual Question Answering}},
  year = {2026},
  eprint = {2601.00269},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.00269},
}
Figure