analiza aiqbench
Meituan reports that its MTFM recommendation model, a unified ranking model across several food delivery businesses, lifted order volume by 2.06% to 6.68% and cut inference cost by 24%. These are results from a single company's production system, not independent benchmarks, and the paper is accepted at KDD 2026. For ML practitioners, the useful details are the techniques: a heterogeneous tokenizer with dynamic masking for multi-scenario feature alignment, hybrid attention for scaling, and a user-level training paradigm.
MTFM: Practical Deployment of Meituan's Unified Recommendation Foundation Model Across Multiple Food Delivery Business Scenarios
Building on MTGR, Meituan proposes MTFM, a unified recommendation foundation model, and for the first time achieves a unified fine-ranking model across multiple major food delivery business scenarios. Addressing three key challenges (scalability, extensibility, and architectural efficiency), MTFM uses a heterogeneous Tokenizer and dynamic masking to achieve feature alignment-free modeling across multiple scenarios, improves scaling capability through hybrid attention and Target Scale-Up, draws on LLM optimization practices to enhance training of deep networks, and uses a User-Level training paradigm and custom operators to keep system overhead under control. Order volume across multiple Meituan food delivery businesses increased by 2.06% to 6.68%, inference costs were reduced by 24%, and the related work has been accepted by KDD 2026.
Meituan builds on MTGR to propose MTFM, a unified foundation model for recommendation. It is the first unified ranking model covering several major Meituan Waimai (food delivery) businesses. To address three challenges, namely scalability, extensibility across scenarios, and architectural efficiency, MTFM uses a heterogeneous Tokenizer and dynamic masking to align multi-scenario features without explicit alignment. It improves the model's scaling capability through hybrid attention and Target Scale-Up, borrows optimization practices from LLMs to improve training of deep networks, and keeps system overhead under control through a User-Level training paradigm and custom operators. Order volume across several Meituan Waimai businesses increased by 2.06% to 6.68%, depending on the business, and inference cost fell by 24%. The related work has been accepted by KDD 2026. Meituan builds on MTGR to propose MTFM, a unified recommendation foundation model. It…
Extras afișat conform termenilor RSS ai publicației · 136 cuvinte · textul integral și figurile sunt pe site-ul publicației
Subiecte conexe
LLM optimization integration for Amazon SageMaker Python SDK
AWS · 6 aug. 2026
Benchmarking small LLM inference on SageMaker AI: G7 vs G5 and G6
AWS · 8 sept. 2026
Semalith v1.4: A Calibrated 184M Safety Classifier Achieving State-of-the-Art Prompt-Injection Detection at 44x Fewer Parameters than Llama-Guard-3-8B
arXiv · 28 iul. 2026
State of the Field: AI for Epistemics and Coordination
LessWrong · 2 oct. 2026