Mitigating Object Hallucination in MLLMs via Data-augmented Phrase-level Alignment

TL;DR

This work addresses object hallucination in multimodal LLMs by introducing Data-augmented Phrase-level Alignment (DPA), which creates correct-hallucinated phrase pairs via generative augmentation and trains models with a phrase-level alignment loss plus a token-wise forward KL regularizer against a frozen reference. The resulting HALVA models mitigate object hallucination across descriptive and discriminative tasks while maintaining or boosting performance on general vision-language benchmarks. Key contributions include the phrase-level loss, careful data augmentation splits, and extensive evaluations across CHAIR, AMBER, MME-Hall, HallusionBench, and non-hallucination benchmarks, demonstrating improved reliability without sacrificing expressiveness. The approach offers practical benefits for deploying reliable, open-source MLLMs in real-world tasks and is accompanied by open-source code and training artifacts.

Abstract

Despite their significant advancements, Multimodal Large Language Models (MLLMs) often generate factually inaccurate information, referred to as hallucination. In this work, we address object hallucinations in MLLMs, where information is generated about an object not present in the input image. We introduce Data-augmented Phrase-level Alignment (DPA), a novel loss which can be applied to instruction-tuned off-the-shelf MLLMs to mitigate hallucinations, while preserving their general vision-language capabilities. To fine-tune MLLMs with DPA, we first generate a set of `hallucinated' and `correct' response pairs through generative data augmentation by selectively altering the ground-truth information of the correct responses at a phrase level. The DPA loss is then used to train MLLMs to reduce the likelihood of hallucinated phrases compared to the correct ones. Our thorough evaluation on various benchmarks confirms the effectiveness of DPA in mitigating hallucination while retaining the out-of-the-box performance of the MLLMs on general tasks. For instance, MLLMs finetuned with DPA, which we refer to as Hallucination Attenuated Language and Vision Assistant (HALVA), improve F1 by up to 13.4% on hallucination visual question-answering and reduce the hallucination rate by up to 4.2% on image description tasks.

Figures (31)

• Figure 1: Examples of object hallucinations.
• Figure 2: (A): A high-level overview comparing the performance of HALVA (the finetuned model with DPA) with existing finetuning methods in mitigating object hallucination, and their ability on general vision-language tasks. (B): Unlike HALVA, the existing finetuning approaches (e.g., HA-DPO and EOS) substantially diverge from their base model (LLaVA-v1.5$_\text{7B}$).
• Figure 3: An example of correct and hallucinated response pairs constructed through our generative data-augmentation. The hallucinated responses are generated by selectively altering the true concepts in the correct response. For instance, we alter 'objects': shirt $\mathrel{ \mkern-4mu\hbox{)}}$ dress, & jeans $\mathrel{ \mkern-4mu\hbox{)}}$ sneakers; 'attributes': white $\mathrel{ \mkern-4mu\hbox{)}}$ black, & blue $\mathrel{ \mkern-4mu\hbox{)}}$ red; 'actions': skateboarding $\mathrel{ \mkern-4mu\hbox{)}}$ rollerblading; and other object-related information such as 'location': skate park $\mathrel{ \mkern-4mu\hbox{)}}$ roller rink. Best viewed in color.
• Figure 4: Overview of our method: Given a vision-language instruction and its correct and hallucinated response pair, the alignment objective ($\mathcal{L}_a$) reduces the log-likelihood of hallucinated tokens compared to the correct ones. Also, a token-wise KL divergence regularizer ($\mathcal{L}_d$) is employed using a reference model ($\pi_{\text{ref}}$), to restrict the divergence of the MLLM ($\pi_\theta$) during DPA training.
• Figure 5: Left: Changes in the model state due to DPA training with varying $\alpha$. Right: Changes is alignment loss before and after training across all training samples. Default $\alpha$ is $0.4$ for HALVA$_\text{7B}$.