paper / edchi / Jun 11
This work introduces a novel framework for real-time generation of LLM-based user interest personas in large-scale video recommendation systems. It addresses challenges of semantic richness, adaptability, and interpretability by creating natural-language personas that balance exploration and exploitation. The system employs a cost-efficient architecture with knowledge distillation, asynchronous inference, and optimized video representations to enable online LLM inference at scale, validated by offline, user, and A/B testing.
large-language-modelsrecommendation-systemsuser-personasonline-inferenceknowledge-distillationpersonalization
“LLMs can enhance recommendation systems by providing world knowledge and reasoning capabilities, addressing limitations of traditional ID-based or offline methods.”
paper / edchi / May 20
Current LLM agent memory evaluations are limited to static, passive retrieval from conversation history — failing to capture the dynamic, continuous learning required in real-world deployments. Evo-Memory addresses this by introducing a streaming benchmark that structures tasks into sequential streams, requiring agents to actively retrieve, adapt, and update memory after each interaction. The paper unifies 10+ memory modules across 10 diverse datasets and proposes ReMem, an action-think-memory refine pipeline designed to enable genuine continual improvement at test time.
llm-agentsmemory-managementai-evalsbenchmarkingtest-time-learningcontinual-learningretrieval-augmented-generation
“Existing LLM memory evaluations focus on static conversational settings and do not assess dynamic accumulation and reuse of experience across evolving task streams.”
paper / edchi / May 16 / failed
paper / edchi / May 16 / failed