SymDPO: Boosting In-Context Learning of Large Multimodal Models with Symbol Demonstration Direct Preference Optimization

TL;DR

Symbol Demonstration Direct Preference Optimization aims to break the traditional paradigm of constructing multimodal demonstrations by using random symbols to replace text answers within instances, and shows that with SymDPO, LMMs can more effectively understand the multimodal context within examples and utilize this knowledge to answer questions better.

Abstract

As language models continue to scale, Large Language Models (LLMs) have exhibited emerging capabilities in In-Context Learning (ICL), enabling them to solve language tasks by prefixing a few in-context demonstrations (ICDs) as context. Inspired by these advancements, researchers have extended these techniques to develop Large Multimodal Models (LMMs) with ICL capabilities. However, existing LMMs face a critical issue: they often fail to effectively leverage the visual context in multimodal demonstrations and instead simply follow textual patterns. This indicates that LMMs do not achieve effective alignment between multimodal demonstrations and model outputs. To address this problem, we propose Symbol Demonstration Direct Preference Optimization (SymDPO). Specifically, SymDPO aims to break the traditional paradigm of constructing multimodal demonstrations by using random symbols to replace text answers within instances. This forces the model to carefully understand the demonstration images and establish a relationship between the images and the symbols to answer questions correctly. We validate the effectiveness of this method on multiple benchmarks, demonstrating that with SymDPO, LMMs can more effectively understand the multimodal context within examples and utilize this knowledge to answer questions better. Code is available at https://github.com/APiaoG/SymDPO.

Figures (6)

• Figure 1: In subfigure (a), an example of visual context overlook is illustrated using OpenFlamingo as a case study. Here, OpenFlamingo Awadalla2023OpenFlamingoAO erroneously generates a response by solely following the textual cues in the demonstration, leading to an inaccurate answer. After applying SymDPO to enhance alignment, OpenFlamingo with SymDPO successfully corrects its response, accurately addressing the question. Subfigure (b) further demonstrates that for OpenFlamingo (OF), replacing images in the demonstration with blank placeholders (OF w/ blank) or omitting images altogether (OF w/o image) surprisingly yields even better performance than the original setup. This result suggests a substantial model dependency on textual context over visual information.
• Figure 2: Comparison of General DPO and SymDPO Formats: General DPO relies solely on standard text for Questions, Answers, Chosen, and Rejected Answers, focusing on text-based training. In contrast, SymDPO replaces textual Answers with symbolized text to boost multimodal understanding, requiring models to interpret both visual and symbolized cues. This approach strengthens the model's ability to reason and decide in complex multimodal contexts.
• Figure 3: Comparison of Symbol Tuning, General DPO, and SymDPO Methods: We optimized OF 3b using three different methods: Symbol Tuning, General DPO, and SymDPO, resulting in three distinct variants. The performance of these variants was visualized using line charts, showcasing the results across four-shot (4, 8, 16, 32) settings on the COCO, VQAv2, and OK-VQA benchmarks.
• Figure 4: Impact of Visual Context Removal on OF and OF+SymDPO Performance.
• Figure 5: Comparison of the Impact of General DPO and SymDPO on LMMs with Varying Data Proportions in the Preference Dataset