Unleashing the Potentials of Likelihood Composition for Multi-modal Language Models

Model fusing has always been an important topic, especially in an era where large language models (LLM) and multi-modal language models (MLM) with different architectures, parameter sizes and training pipelines, are being created all the time. In this work, we propose a post-hoc framework, aiming at fusing heterogeneous models off-the-shell, which we call $\mathit{\text{likelihood}}\mathit{\text{composition}}$, and the basic idea is to compose multiple models’ likelihood distribution when doing a multi-choice visual-question-answering task. Here the core concept, $\mathit{\text{likelihood}}$, is actually the log-probability of the candidate answer. In $\mathit{\text{likelihood}}\mathit{\text{composition}}$, we introduce some basic operations: $\mathit{\text{debias}}$, $\mathit{\text{highlight}}$, $\mathit{\text{majority-vote}}$ and $\mathit{\text{ensemble}}$. By combining (composing) these basic elements, we get the mixed composition methods: $\mathit{\text{mix-composition}}$. Through conducting comprehensive experiments on 9 VQA datasets and 10 MLMs, we prove the effectiveness of $\mathit{\text{mix-composition}}$ compared with simple $\mathit{\text{ensemble}}$ or $\mathit{\text{majority-vote}}$ methods. In this framework, people can propose new basic composition methods and combine them to get the new mixed composition methods. We hope our proposed $\mathit{\text{likelihood}}\mathit{\text{composition}}$ can provide a new perspective of fusing heterogeneous models and inspire the exploration under this framework.