Reducing Hallucinations in Vision-Language Models via Latent Space Steering

TL;DR

This work identifies vision feature instability as a core contributor to hallucinations in vision-language models and introduces Visual and Textual Intervention (VTI), a training-free, test-time latent-space steering method. By computing stable-direction vectors from perturbation-based vision and non-hallucinated captions, and applying these directions to both the vision encoder and the text decoder, VTI reduces hallucinations without extra training or cost. Empirical evaluations across POPE, CHAIR, and MMHAL-Bench show consistent improvements over decoding-based baselines on multiple LVLM backbones, highlighting the method's robustness and generality. The study emphasizes latent-space stability as a critical lever for reliable multimodal reasoning in LVLMs.

Abstract

Hallucination poses a challenge to the deployment of large vision-language models (LVLMs) in applications. Unlike in large language models (LLMs), hallucination in LVLMs often arises from misalignments between visual inputs and textual outputs. This paper investigates the underlying mechanisms of hallucination, focusing on the unique structure of LVLMs that distinguishes them from large language models (LLMs). We identify that hallucinations often arise from the sensitivity of text decoders to vision inputs, a natural phenomenon when image encoders and text decoders are pre-trained separately. Inspired by this, we introduce Visual and Textual Intervention (VTI), a novel technique designed to reduce hallucinations by steering latent space representations during inference to enhance the stability of vision features. As a task-agnostic test-time intervention, VTI can be easily applied to any problem without additional cost. Extensive experiments demonstrate that it can effectively reduce hallucinations and outperform baseline methods across multiple metrics, highlighting the critical role of vision feature stability in LVLMs.

Figures (9)

• Figure 1: Illustration of the effect of our proposed method, VTI, using LLaVA-1.5. Hallucinated contents generated by the original model are marked in red. In contrast, VTI results in less hallucination across different categories of questions. Examples are obtained from MMHAL-Bench sun2023aligning and CHAIR chair
• Figure 2: (Left) Distribution of vision feature stability when different types of noise are injected into the raw images. The $x$-axis represents the variance of features across 50 perturbations, and the $y$-axis represents the frequency. (Middle) Illustration of the correlation between object hallucination and vision feature stability. Averaging vision features across multiple perturbed images reduces hallucination as the number of perturbations averaged increases. (Right) Noise alone tends to increase hallucination, suggesting that the reduction of hallucination is not due to the noise itself but to the averaging process across perturbations.
• Figure 3: Overview of the proposed algorithm visual and textual test-time intervention (VTI). Given an example set $\{(v_i, x_i, \tilde{x}_i)\}_{i=1}^N$ where $v_i$ is the vision input and $(x_i,\tilde{x}_i)$ is paired captions with and without hallucination, VTI first runs the model on each query $(v_i, x_i, \tilde{x}_i)$ and records all hidden states. It then computes the shifting vectors $d_{l,t}^\text{vision}$ and $d_{l,t}^\text{text}$ for all layer $l$ and token $t$ according to \ref{['sec:method']}. During inference, the vectors are subsequently added to every layer of the vision encoder and text decoder, respectively, when processing a new query. Notice that the vectors are task- and dataset-agnostic, i.e., they are pre-computed using a few samples from one specific task and dataset, and fixed unchanged throughout the entire experiments in our paper.
• Figure 4: Detailed performance of different methods with LLaVA-1.5 as the backbone on the eight categories in MMHAL-BENCH sun2023aligning, where "Overall" indicates the averaged performance across all categories. A higher score indicates that the generated response contains fewer hallucinations and more information.
• Figure 5: Feature stability is improved across different types of image corruptions with vision intervention.