This episode explores the evolution of recommender systems, moving beyond traditional collaborative filtering to agentic approaches powered by Large Language Models (LLMs). It discusses how LLMs transform recommenders from simple ranking tools into task-oriented agents capable of complex, conversational interactions, while also addressing critical aspects like trustworthiness and emerging risks such as hallucination.
Summarized by Podsumo
Recommender systems are evolving from merely ranking items to performing complex, multi-constraint tasks for users, driven by LLMs and language agents.
Large Language Models enhance recommender systems by enabling natural, conversational interactions and augmenting personalization through external knowledge and multimodal signal processing.
The discussion covers five dimensions of trustworthy AI (generalizability, robustness, privacy, explainability, fairness) and new risks introduced by LLMs, such as hallucination and context drift.
Agents can replace classic models, augment existing recommender systems (e.g., data augmentation), or be used for simulation and evaluation of user and system behavior.
A key framework for agents includes an LLM as the brain, input/output spaces, tools to enhance capabilities, and various memory types (working, episodic, semantic, procedural) crucial for engaging, human-like conversations.
"Now, instead of do this ranking for me, we can say with large language models and language agents where we are heading to has been transformed to please do these tasks for me."
— Yashar Delju
"The memory aspect is what has made this system a favorite by users because when you're discussing just like human human conversation just like we're discussing with human human we would like some that can remember your goals and priorities."
— Yashar Delju
"Their creativity, looking at the problem, maybe from a new perspective, is something that if it's used with care, with care could be nice."
— Yashar Delju