Researchers have developed TGR-EvoDAG, a novel multi-agent framework designed for the autonomous evolution of code in Generative Recommendation (GR) systems. This Directed Acyclic Graph (DAG)-based approach allows for the maintenance of multiple evolution branches and the merging of complementary innovations, addressing limitations in existing methods such as cost blindness and short-sighted evaluation. TGR-EvoDAG optimizes GPU-hour budgets through stage-aware cost planning and navigates temporary metric regressions with active downstream co-adaptation. The framework has demonstrated significant improvements, including a 37.4% increase in offline metrics and a 19.7% rise in simulated user satisfaction on Amazon datasets, and a 12.16% improvement in Hit@1 and a 0.40% increase in online click-through rate for Tencent's media platform. AI
IMPACT Enables autonomous code evolution for generative recommendation systems, potentially improving efficiency and performance in real-world applications.
RANK_REASON The item is an arXiv paper detailing a new framework and its performance on datasets and in an industrial deployment. [lever_c_demoted from research: ic=1 ai=1.0]
- Agent4Rec
- directed acyclic graph
- Generative Recommendation
- graphics processing unit
- Hugging Face
- large language model
- Relative strength index
- Tencent
- TGR-EvoDAG
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
