SAERec uses sparse autoencoders to extract fine-grained intents from LLMs for better recs

SAERec tackles a key flaw in existing intent-based recommender systems: they derive intents directly from user sequences via clustering or prototype learning, making them sensitive to sequence quality, requiring a preset number of intents, and lacking explicit semantic grounding. As a result, the intent sets are often incomplete and coarse, limiting recommendation accuracy and interpretability.

To solve this, SAERec leverages a textual corpus (e.g., item descriptions) as high-information-density evidence for intent construction. It first uses a sparse autoencoder (SAE) to disentangle LLM text embeddings, isolating intent-related semantics from noise and extracting a comprehensive set of fine-grained, interpretable intents. For each user, it then retrieves relevant intents—both personal (current interests) and public (general patterns like quality or price)—as priors to guide recommendation.

These priors are integrated into sequence modeling via a multi-branch attention mechanism that captures temporal dependencies while injecting personal and public intent signals. An adaptive fusion layer constructs the final user representation. Extensive experiments on public datasets show SAERec consistently outperforms state-of-the-art baselines and offers human-understandable explanations for its recommendations.