Highlights of Meituan’s KDD’26 Papers and the Champion Strategies for the DataAgents Track

KDD (ACM SIGKDD Conference on Knowledge Discovery and Data Mining) is a top‑tier international conference in data mining with an acceptance rate of roughly 15‑20% (CCF‑A venue).

MTFM: A Scalable and Alignment‑free Foundation Model for Industrial Recommendation

Existing cross‑domain (CDR) and multi‑scenario (MSR) recommendation methods require heavy resources and strict input alignment, limiting scalability. MTFM replaces alignment with heterogeneous tokenization of cross‑domain data, allowing a Transformer‑based model to learn multi‑scenario knowledge without pre‑alignment. A user‑level sample aggregation mechanism reduces the total number of training instances, increasing throughput. Grouped‑Query Attention and a customized Hybrid Target Attention lower memory consumption and compute cost. System‑level optimizations (kernel fusion, elimination of CPU‑GPU blocking) further accelerate training and inference. Offline and online experiments in food‑delivery scenarios show significant performance gains, enabling a single recommendation foundation model to replace independent ranking models across major business scenes.

CDRRM: Contrast‑Driven Rubric Generation for Reliable and Interpretable Reward Modeling

Traditional reward models are opaque and rely on expensive labeling. CDRRM introduces a “contrast‑aggregate” pipeline: first contrast good and bad responses to locate key differences, then aggregate those differences into concise, task‑specific rubrics that guide the evaluation model. Experiments on three benchmarks achieve state‑of‑the‑art results, reduce verbosity and positional biases, and require only 3 k samples for an un‑fine‑tuned model to surpass a fully fine‑tuned baseline, demonstrating efficiency and interpretability.

LocalSearchBench: Benchmarking Agentic Search in Real‑World Local Life Services

To address the lack of evaluation data for AI‑driven search in local life services, the authors construct LocalSearchBench, covering nine Chinese cities, six service categories, and 900 multi‑hop QA tasks. The benchmark provides the interactive environment LocalPlayground and a merchant retrieval tool LocalRAG. Evaluation of 16 mainstream large‑language inference models reveals poor information completeness and credibility, and frequent failures in tool use and multi‑hop reasoning, highlighting research directions for agentic search.

Deterministic‑Allocation and Anonymous Joint Advertising

Joint ad auctions require both anonymity (outcome independent of bidder identity) and deterministic allocation (unique result for identical inputs). Existing algorithms cannot satisfy both, causing unfair distribution and incentive misalignment. JTransNet integrates anonymity and deterministic allocation via a differentiable NeuralSort ranking method, forming an end‑to‑end data‑driven model that resolves multi‑party traffic allocation and charging. Offline and online experiments report significant revenue gains, and the model has been fully deployed in Meituan Retail.

UME: A Unified Meta‑Generalization Framework for Cross‑Domain ETA

Accurate checkout‑page ETA prediction faces three challenges: (1) zero‑shot prediction for unseen markets, (2) heterogeneous feature spaces across domains, and (3) separate modeling of mature and cold‑start markets. UME introduces a dual‑branch network and a hyper‑network‑based meta‑learning mechanism that dynamically modulates feature gates, expert attention, and predictions using domain‑level knowledge and instance‑level context. Knowledge distillation bridges feature‑gap information loss. Offline and online experiments show UME outperforms prior methods.

GRAD: Generative Large‑Scale Pre‑trained Models for Automated Ad Bidding Optimization

Automated bidding must balance overall performance, diverse advertiser goals, and real‑world constraints (e.g., CPM, ROI). Conditional generative models (Transformers, diffusion) can generate bidding trajectories but suffer from distribution shift, limited action‑space exploration, and constraint violations. GRAD addresses these issues with a mixed‑expert action module for diverse bidding behaviors and a causal Transformer that optimizes under constraints.

HMAF: A Hierarchical Multi‑Slot GD‑RTB Allocation Framework

Guaranteeing both guaranteed‑delivery (GD) contracts and real‑time bidding (RTB) auctions in multi‑slot environments requires coordinated allocation. Existing solutions either decouple GD and RTB optimization or rely on heuristic priority rules, failing to balance short‑term revenue and long‑term contract fulfillment. HMAF adopts a “plan‑calibrate‑execute” paradigm: offline GD resource planning, dynamic calibration of GD vs. RTB competition intensity, and online list‑level ranking for each slot, unifying offline constraint optimization with online decision making.

MTGenRec: An Efficient Distributed Training System for Generative Recommendation Models

PyTorch’s support for large‑scale sparse embedding training is limited. MTGenRec builds a unified sparse‑dense training framework on the PyTorch ecosystem. Key techniques include:

Dynamic hash tables replace static ID tables, handling dynamic ID lifecycles.

Automatic table merging, ID deduplication, and variable‑length sequence load balancing improve efficiency.

Checkpoint‑resume, mixed‑precision training, gradient accumulation, and operator fusion further accelerate training.

Experiments show 1.6‑2.4× speedup over TorchRec without accuracy loss and near‑linear scaling from 8 to 128 GPUs. MTGenRec is deployed in multiple core Meituan scenarios.

KDD Cup 2026 DataAgents Track – Champion Strategy

The champion solution builds on the internal “Agent Harness” from the “问点仔” project, creating a runtime that autonomously explores heterogeneous file types (databases, PDFs, JSON, charts, videos) and automatically invokes sub‑agents for SQL, Python, video understanding, and unstructured document ETL. Mechanisms for error feedback, timeout control, and automatic retries improve stability on long‑chain tasks. The open‑source code is available at https://github.com/zhezh/kddcup2026_champion.