Duet: Joint Exploration of User–Item Profiles

ICLR 2026 Conference Submission23981 Authors

Keywords: recommendation, large language model, representation learning

Abstract: Traditional recommendation systems represent users and items as hidden vectors, learning to align them in a shared latent space for relevance estimation. With the advent of large language models (LLMs), we advocate a shift *from vectors to text*: representing users and items as *textual profiles* and aligning them in a shared semantic space. Textual profiles are directly compatible with LLMs and offer interpretability for downstream agentic systems. A key challenge, however, is that the optimal profile format is unknown, and handcrafted templates often misalign with task objectives. We propose **Duet**, a framework for *joint exploration of user–item profile generation in text*. The framework operates in three stages. First, raw histories and metadata are distilled into simple *cues* that capture minimal but informative signals. Second, during a single sequence-to-sequence inference pass, these cues are expanded into richer prompts and then into textual profiles, allowing for the exploration of multiple formats rather than a single creation. Finally, profiles are optimized jointly via reinforcement learning, where downstream recommendation performance provides feedback to refine and align them. Experiments on three real-world datasets demonstrate that **Duet** outperforms strong baselines, validating the effectiveness of joint textual profile alignment and the utility of prompt-driven exploration. Project page: [https://duetreview.github.io/](https://duetreview.github.io/).

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 23981