Source-linked AI summary
VIB-Probe: Detecting and Mitigating Hallucinations in Vision-Language Models via Variational Information Bottleneck
Feiran Zhang, Yixin Wu, Zhenghua Wang, Xiaohua Wang, Changze Lv, Xuanjing Huang, Xiaoqing Zheng
TL;DR
VLM hallucination detection often overlooks internal mechanisms and must separate predictive signals from noise in high-dimensional attention states. VIB-Probe uses a variational information bottleneck to compress attention features, then identifies and suppresses hallucination-sensitive heads at inference time. The framework reports state-of-the-art detection and mitigation across diverse benchmarks, while remaining limited to white-box transformer-based VLMs with explicit attention.
Problem
Existing hallucination detectors primarily rely on output statistics or external verification, overlooking internal mechanisms and facing noisy, high-dimensional attention states.
Method
VIB-Probe compresses multi-layer, multi-head attention outputs into a predictive latent representation and uses gradient attribution to suppress hallucination-sensitive heads during decoding.
Results
VIB-Probe achieves state-of-the-art hallucination detection and mitigation across diverse benchmarks, perturbations, and VLM architectures.
Takeaways & Limitations
The framework supports lightweight inference-time mitigation by identifying and down-weighting a small set of hallucination-sensitive attention heads when risk is high.
Takeaways & Limitations
The method is evaluated primarily on transformer-based VLMs with standard attention and requires internal representations, restricting it to white-box settings.
Abstract
from arXiv · showhide
Vision-Language Models (VLMs) have demonstrated remarkable progress in multimodal tasks, but remain susceptible to hallucinations, where generated text deviates from the underlying visual content. Existing hallucination detection methods primarily rely on output logits or external verification tools, often overlooking their internal mechanisms. In this work, we investigate the outputs of internal attention heads, postulating that specific heads carry the primary signals for truthful generation.However, directly probing these high-dimensional states is challenging due to the entanglement of visual-linguistic syntax and noise. To address this, we propose VIB-Probe, a novel hallucination detection and mitigation framework leveraging the Variational Information Bottleneck (VIB) theory. Our method extracts discriminative patterns across layers and heads while filtering out semantic nuisances through the information bottleneck principle. Furthermore, by leveraging the gradients of our VIB probe, we identify attention heads with strong causal influence on hallucinations and introduce an inference-time intervention strategy for hallucination mitigation. Extensive experiments across diverse benchmarks demonstrate that VIB-Probe significantly outperforms existing baselines in both settings. Our code will be made publicly available.
1 Introduction
VLM hallucinations undermine reliable multimodal reasoning, while existing detectors often use limited output statistics. VIB-Probe instead compresses multi-layer, multi-head attention signals for detection and uses attribution to suppress hallucination-sensitive heads during inference.
- VLM hallucinations produce descriptions unfaithful to image objects or relations, undermining reliability in high-stakes multimodal applications.
- Existing detectors rely mainly on surface-level confidence features that may generalize poorly across architectures.These methods exploit only a narrow portion of the model’s internal dynamics.
- VIB-Probe distills multi-layer, multi-head attention outputs into a compact representation that retains hallucination-predictive information while suppressing noise.The framework applies Variational Information Bottleneck theory to internal attention features.
- Gradient-based attribution identifies hallucination-sensitive attention heads, which are selectively suppressed during decoding when hallucination risk exceeds a threshold.This inference-time strategy connects hallucination detection with mitigation without requiring retraining.
- VIB-Probe achieves state-of-the-art performance in hallucination detection and mitigation across diverse benchmarks, perturbations, and architectures.
2 Related Work
Prior work addresses VLM hallucinations through output statistics, attention analysis, training, verification, and inference-time interventions. VIB-Probe combines attention-head probing with an information bottleneck and gradient-based targeting for training-free intervention.
- Hallucinations in VLMs: VLM hallucinations are commonly categorized as object, attribute, or relational failures arising from visual-grounding errors.
- Hallucination Detection: Early detection methods use shallow confidence or entropy statistics, while newer approaches probe attention patterns to improve discrimination.
- Hallucination Mitigation: Mitigation methods include training-based robustness improvements, post-generation external verification, and inference-time attention interventions.
- Hallucination Mitigation: VIB-Probe uses gradient-based attribution to target hallucination-sensitive attention heads for training-free inference-time intervention.
- Information Bottleneck: The Information Bottleneck principle regularizes representations by compressing inputs to discard irrelevant features.
- Information Bottleneck: Variational Information Bottleneck operationalizes this trade-off through a variational approximation inspired by variational autoencoders.
3 Method
VIB-Probe detects hallucinations from internal multi-layer, multi-head attention outputs by compressing them into predictive bottleneck representations. It then uses probe gradients to select influential heads for threshold-triggered inference-time suppression.
- Attention head outputs: The method aggregates pre-projection attention head outputs across Transformer layers and heads into a tensor representing internal multimodal processing.For each decoding step, the tensor has dimensions L×H×d_h.
- Variational information bottleneck: VIB-Probe applies an information bottleneck that retains label-predictive information while suppressing irrelevant features and semantic nuisances.The objective balances prediction against compression using a variational approximation and a prior distribution.
- VIB detector: The detector encodes the attention tensor into a latent representation, samples during training, and uses its mean representation for deterministic inference-time hallucination-risk prediction.A classification layer converts the latent representation into a risk logit and sigmoid probability.
- Training objective: Training combines binary cross-entropy with KL regularization, while β controls the trade-off between prediction accuracy and representation compression.The KL term regularizes the approximate posterior toward a standard normal prior.
- Hallucination mitigation: At each decoding step, the model intervenes only when the risk logit exceeds a threshold, modifying attention heads before regenerating the token.The intervention is designed to translate detected hallucination risk into model control during inference.
- Gradient-based attribution and head selection: Head importance is computed from the magnitude of each head’s risk-logit sensitivity, and the most influential heads are selected for targeted suppression.The method obtains sensitivity through gradients with respect to each attention head output.
4 Experiments
VIB-Probe is evaluated across discriminative and generative hallucination benchmarks, four VLMs, transfer settings, robustness perturbations, and mitigation tasks. It consistently outperforms existing methods, with stronger gains on generative tasks and stable transfer under distribution shifts.
- Benchmarks: Experiments cover POPE, AMBER, M-HalDetect, and COCO-Caption across MiniGPT-4, LLaVA-v1.5-7B, LLaVA-v1.6-Mistral-7B, and Qwen2.5-VL-7B-Instruct.Detection uses AUROC and AUPRC across discriminative and generative settings.
- Detection performance: VIB-Probe consistently outperforms existing state-of-the-art methods across the four evaluated VLMs.The comparison includes uncertainty-based, probing, MetaToken, and DHCP baselines.
- Detection performance: +1.20% on discriminative benchmarks and +2.84% on generative tasks, with the larger gain occurring in challenging free-form text.The results indicate stronger detection capability on generative hallucinations.
- Transferability: 32.4% decline on M-HalDetect affects RepProbing under domain shift, whereas VIB-Probe exhibits stable and stronger transferability.The transfer experiments train on POPE-Popular and evaluate across discriminative and generative tasks.
- Robustness: VIB-Probe maintains high detection accuracy under rotation, Gaussian blur, and brightness perturbations applied to POPE and COCO-Caption images.Perturbations are used only for evaluation while ground-truth labels remain valid.
- Ablation studies: Removing the KL loss reduces performance to a level comparable to RepProbing, demonstrating the importance of the Information Bottleneck constraint.The ablation retains the encoder-decoder structure and optimizes only the BCE loss.
- Ablation studies: Using only a small subset of layers degrades performance, while deeper layers outperform shallower layers, especially on M-HalDetect.The experiment compares layer selections against using attention heads from all VLM layers.
5 Conclusion
VIB-Probe addresses hallucinations arising progressively from internal attention dynamics by distilling multi-head, multi-layer signals through a Variational Information Bottleneck. It also enables lightweight inference-time mitigation by identifying and down-weighting hallucination-sensitive heads, with experiments reporting state-of-the-art detection and mitigation performance.
- VIB-Probe targets hallucinations that emerge progressively from internal attention dynamics rather than solely from the final output.
- The framework distills high-dimensional attention outputs across all layers into compact representations that isolate hallucination-related signals from task-irrelevant noise.
- VIB-Probe identifies and down-weights a small set of hallucination-sensitive heads during inference when hallucination risk is high.
- Experiments across diverse architectures and benchmarks demonstrate state-of-the-art performance in both hallucination detection and mitigation.
Limitations
The study focuses on transformer-based VLMs with standard attention mechanisms and requires access to internal representations, restricting VIB-Probe to white-box settings.
- VIB-Probe has been evaluated primarily on transformer-based VLMs with standard attention mechanisms.Its applicability to alternative multimodal architectures without explicit attention structures remains unexplored.
- Access to internal representations and attention outputs restricts the method to white-box settings.
A Models and Baselines
The paper evaluates several VLM backbones and compares uncertainty, hidden-state probing, decoding, and visual-grounding baselines for hallucination detection or mitigation. These baselines span sentence-level uncertainty measures, hidden-state classifiers, beam search, and training-free interventions.
- Models: MiniGPT-4, LLaVA-v1.5-7B, LLaVA-v1.6-Mistral-7B, and Qwen2.5-VL-7B-Instruct provide the evaluated VLM backbones.The first three listed backbones use 32 layers and 32 attention heads, while Qwen2.5-VL-7B-Instruct uses 28 layers and 28 attention heads.
- Detection baselines: AvgProb measures sentence-level uncertainty from the mean negative log-probability of generated tokens.Higher AvgProb indicates lower likelihood assigned to produced tokens and greater sentence-level uncertainty.
- Detection baselines: AvgEnt measures sentence-level uncertainty by averaging token-level predictive entropies across the vocabulary distribution.Higher AvgEnt indicates more diffuse predictive distributions across tokens.
- Detection baselines: RepProbing uses a lightweight classifier trained on last-layer decoder hidden states to estimate hallucination risk.
- Decoding baselines: Beam search generates sequences by repeatedly retaining the top-B partial hypotheses according to cumulative log-probability.
- Mitigation baselines: PAI boosts attention to image tokens and subtracts text-only logits from multimodal logits to reduce language-only bias.VCD instead contrasts outputs from original and distorted images to suppress statistical biases and unimodal language priors.
B.1 Hallucination Detection
The hallucination-detection experiments construct labeled responses from benchmark images, questions, and ground truths, using separate dataset splits and task-specific evaluation settings. They cover discriminative, span-based, and sentence-based hallucination detection.
- POPE and AMBER samples are converted into responses containing or avoiding hallucinations according to ground-truth labels.
- POPE uses training and validation splits across popular, random, and adversarial subsets, while AMBER uses a curated 5,000-sample subset.Samples associated with the same image are kept from overlapping between training and validation sets.
- M-HalDetect uses an 80:20 training-validation split and reports span-based hallucination detection results.
- COCO-Caption responses are generated with LLaVA-v1.5-7B, annotated using official COCO 2014 Val annotations, and evaluated for sentence-based hallucination detection.
B.2 Model Architecture
VIB-Probe reduces original attention output feature vectors with a multilayer encoder, then processes them through residual blocks and a linear decoder.
- The encoder is a 3-layer MLP with dimensions (1024, 512, 256) for reducing original attention output feature vectors.
- Two residual blocks further process the reduced features, while the decoder uses a single linear layer.
- GELU activation and LayerNorm are applied throughout the network.
B.3 Hallucination Mitigation
The experiments evaluate object hallucinations by comparing generated captions with ground-truth object annotations and intervene on the most important attention heads.
- Object hallucinations are evaluated using CHAIR metrics, which compare generated captions against ground-truth object annotations.
- CHAIRi measures the proportion of hallucinated object mentions among all generated object mentions.
- CHAIRs measures the percentage of captions containing at least one hallucinated object.
- Mitigation intervenes on attention heads ranked in the top 5% by importance scores.
- The intervention threshold uses average training-set logit values, with suppression strength λ set to 0.001.