HALVA: Hallucination Attenuated Language and Vision Assistant

Recent advancements in Large Language Models (LLMs) have laid the foundation for the development of highly capable multimodal LLMs (MLLMs) like Gemini. MLLMs can process additional modalities, such as images or videos, while retaining language understanding and generation capabilities. Despite the impressive performance of MLLMs across a variety of tasks, the issue of object hallucination presents a significant challenge to their widespread adoption. Object hallucination refers to generated language that includes descriptions of objects or their attributes that are not present in or cannot be verified by the given input.

Our method consists of two key steps: (1) generative data augmentation and (2) contrastive tuning. For a given pair of vision-language instructions and a corresponding correct answer, generative data augmentation is applied to obtain hallucinated responses. We selectively alter ground-truth objects and object-related attributes from the correct answer to introduce hallucinated concepts that are not present in the input images.

A contrastive loss is then calculated between a pair of factual and hallucinated tokens. Our objective is to minimize the likelihood of generating hallucinated tokens and correspondingly, maximize the likelihood of generating factual tokens. We train the MLLM with a KL-divergence regularizer that ensures that the MLLM retains its original performance in general vision-language tasks by preventing it from diverging from the base model. We refer to MLLMs trained with the contrastive tuning framework as Hallucination Attenuated Language and Vision Assistants (HALVA).

We use LLaVA-v1.5, a widely used open-sourced MLLM, as our base model and train it using our contrastive tuning framework (HALVA). We then evaluate its performance on object hallucination mitigation and general visual question answering tasks (VQA) against fine-tuning–based approaches, HA-DPO and EOS. We consider LLaVA-v1.5 as the lower bound and GPT-4V as a strong reference point given its performance on standard benchmarks.

We use the AMBER benchmark and Caption Hallucination Assessment with Image Relevance (CHAIR) metric to evaluate MLLM performance on image description tasks, assessing both hallucination rate and the level of detail of their generated image descriptions. The latter aspect is quantified by calculating the percentage of ground-truth objects present in the image that are accurately captured in the model’s output. Our goal is to mitigate hallucinations while retaining or improving the richness of image descriptions. As shown in the left plot below, HALVA captures more ground-truth objects while hallucinating less than HA-DPO. Moreover, while EOS achieves a slightly lower hallucination rate, it degrades the level of detail in the image descriptions, performing worse than HALVA.

We also use the F1-score to compare the performance of MLLMs on visual question answering tasks using the AMBER benchmark for object hallucination and TextVQA benchmark for general vision language accuracy. As shown in the right plot below, both HA-DPO and EOS underperform HALVA in mitigating object hallucination and even deteriorate general vision-language abilities compared to the base model.

When evaluating performance on image description tasks, we use CHAIR to measure hallucination at two levels: instance-level and sentence-level. In this task, HALVA is prompted with “Describe the image in detail”, which triggers it to generate detailed image descriptions. On the MS COCO dataset, HALVA substantially reduces hallucination compared to the base variants. For instance, compared to LLaVA-v1.5, the 7B (HALVA-7B) and 13B (HALVA-13B) variants reduce instance-level hallucinations from 15.4 to 11.7 points and from 13 to 12.8 points, respectively. On the AMBER benchmark, the frequency of hallucinated objects in image descriptions is also captured by CHAIR, where HALVA-7B and HALVA-13B reduce hallucinations from 7.8 and 6.6 points to 6.6 and 6.4 points, respectively.

We evaluate HALVA on discriminative tasks using the MME benchmark. Specifically, we utilize the hallucination subset of MME, which consists of four object-related subtasks — existence, count, position, and color — referred to as MME-Hall. The results presented below demonstrate that HALVA substantially improves performance compared to the base model, LLaVA-v1.5. For instance, HALVA-13B achieves a score of 675 out of 800, resulting in a performance gain of 31.7 points compared to the base model.

To further stress-test our method on other forms of vision-language hallucinations that are not restricted to objects and may occur due to visual illusions, we evaluate performance using the HallusionBench benchmark. Our results demonstrate that HALVA also directly benefits other forms of vision-language hallucinations. HALVA 7B and 13B variants improve overall accuracy by 1.86% and 2.16%, respectively, compared to their base models.

To mitigate object hallucination in MLLMs, we introduce data augmented contrastive tuning. Our proposed method is effective in mitigating object hallucinations and beyond while retaining or even improving their performance on general vision-language tasks. Moreover, the proposed contrastive tuning is simple, fast, and requires minimal training with no additional overhead at inference. We believe that our method may have applications in other areas as well. For example, it might be adapted to mitigate bias and harmful language generation, among others.

We gratefully acknowledge the contributions of our co-authors, Ali Etemad, Ahmad Beirami, Sercan Arik and Tomas Pfister. Special thanks to Tom Small for creating the animated figure in this blog post.