TLDR: OMNI DPO is a novel preference optimization framework designed to mitigate hallucinations in Omni-modal Large Language Models (OLLMs). It addresses the issue of textual dominance and insufficient audio-visual alignment by constructing text-preference and multimodal-preference sample pairs. This approach encourages OLLMs to better integrate visual and auditory information, leading to a significant reduction in hallucinations and improved reasoning capabilities, as demonstrated through experiments on Qwen2.5-Omni and MiniCPM-o-2.6.
The rapid advancements in Omni-modal Large Language Models (OLLMs) have opened up new possibilities for understanding and generating content from various inputs like audio, video, and text. These models are capable of complex tasks, from answering questions about videos to describing audio-visual scenes. However, a significant challenge persists: hallucination. This is when models produce outputs that don’t accurately reflect the given visual or auditory information, leading to inconsistent or fabricated content. This issue is particularly concerning in critical applications such as medical video analysis or autonomous driving, where accuracy is paramount.
A key problem identified in OLLMs is that the text modality often dominates, causing models to rely too heavily on textual cues and neglect visual and audio information. This is similar to issues seen in bimodal (text-image) models. Furthermore, omni-modal scenarios introduce unique challenges because many existing models align visual or auditory modalities with text independently during training, often ignoring the natural connections between video and its corresponding audio. This oversight can lead to hallucinations when reasoning requires interpreting subtle audio cues embedded within video content.
To tackle these challenges, researchers have proposed OMNI DPO, a new preference-alignment framework specifically designed to reduce hallucinations in OLLMs. OMNI DPO employs two main strategies. First, it constructs “text-preference sample pairs” to improve the model’s understanding of how audio and video interact. Second, it builds “multimodal-preference sample pairs” to strengthen the model’s focus on visual and auditory information. By addressing both the dominance of textual priors and the insufficient alignment between audio and visual modalities, OMNI DPO aims to enhance multimodal grounding and significantly reduce hallucinations.
The OMNI DPO framework extends the concept of Direct Preference Optimization (DPO) to omni-modal settings. It introduces conditional preference learning for each modality. To facilitate this, the team created a new dataset called OMNI DPO-10k. This dataset contains about 10,000 samples, each with a question and paired textual preferences, as well as audio and video modality preferences. For instance, to create positive textual preference samples, audio descriptions are extracted and combined with video to generate an answer that reflects audio understanding. For negative samples, the same video is used without its audio, leading to an answer that lacks audio comprehension. This helps the model learn to value genuine audio cues. To further address the issue of dominant textual priors, noisy versions of video and audio inputs are introduced to create modality preference pairs, encouraging the model to pay closer attention to visual and auditory signals rather than relying solely on text.
Experiments were conducted on two prominent OLLMs, Qwen2.5-Omni and MiniCPM-o-2.6. The results showed that OMNI DPO not only effectively reduces omni-modal hallucinations but also significantly improves the models’ reasoning capabilities across different modalities. For example, on the AVHBench benchmark, OMNI DPO led to average F1-score gains of 5.82% for Qwen2.5-Omni and 2.64% for MiniCPM-o-2.6. On the CMM benchmark, it consistently improved both Perception Accuracy (PA) and Hallucination Resistance (HR), with notable gains in HR, indicating its effectiveness in reducing false affirmations of non-existent multimodal information.
An ablation study revealed that optimizing preferences for individual modalities (audio-only or video-only) had positive effects, and combining both led to even greater robustness against hallucination. Interestingly, optimizing for only one modality could sometimes negatively impact the other, highlighting the importance of a joint optimization approach. Beyond hallucination mitigation, OMNI DPO also enhanced general reasoning capabilities on benchmarks like MMAU and MMMU, suggesting that encouraging better attention to audio and visual inputs through multimodal preference optimization has broader benefits.
Also Read:
- Ensuring Language Models Reason for the Right Reasons
- SATQuest: A New Approach to Evaluating and Enhancing AI’s Logical Reasoning
This research marks a significant step forward in making OLLMs more reliable and trustworthy, especially in applications where factual precision is critical. For more technical details, you can refer to the full research paper available at https://arxiv.org/pdf/2509.00723.
