Limitations

While we argue and demonstrate in this work that using a UEM is an efficient way to encode long user histories with easier extensions to multimodal inputs, we acknowledge that text prompting can be further optimized, by using text-to-text prompt compression models. These trade-offs could be further studied. The simplicity of the UEM architecture leaves a lot of headroom as demonstrated by LongT5 baselines in Table 1. Our presentations for U are using generic semantic embeddings with the use of SentenceT5 (Ni et al., 2022), these can be further improved with the use of domain specific embeddings. Our experiments are using LMs that are <1B parameters, which are usually considered smaller family of LLMs. It would be a good future direction to consider larger models with parameter efficient tuning techniques. Furthermore, our research has primarily focused on preference understanding, and hasnt been tested on tasks extending to areas such as rating prediction or item recommendati

We expect our conclusions here are likely apply to these tasks. We plan to address these limitations and pursue these avenues in our future research efforts.

Ethics Statement The datasets and models utilized in this study are based on publicly available and open-source resources. While we acknowledge the inherent ethical considerations associated with language models, we do not anticipate any additional ethical concerns arising from the datasets and models developed in the course of this research.