KDD (ACM SIGKDD Conference on Knowledge Discovery and Data Mining) is the most influential top international academic conference in the field of data mining and knowledge discovery. Known for its strict paper acceptance standards and profound academic influence, KDD is an important platform for promoting data-driven research and application innovation. The conference provides a high-level forum for the academic and industrial communities to exchange cutting-edge results. The paper acceptance rate is usually around 15%-20%. It is a top CCF-A conference in the computer field.
This article selects and shares 8 papers by the Meituan technical team that were included in KDD 2026. These papers cover recommendation large models, generation and reward modeling frameworks, agent search, Transformer framework, meta-generalization framework and other technical fields.
🏆 In addition, Dianping’s Technology Department also won the championship and third runner-up in the 2026 KDD Cup Complex Data Analysis Data Agents international competition. This article introduces the team’s competition ideas and problem-solving strategies. I hope the above content can be helpful or inspiring to everyone.
01 MTFM: A Scalable and Alignment-free Foundation Model for Industrial Recommendation in Meituan
Paper download：PDF
Introduction to the paper: Industrial recommender systems usually involve multiple scenarios, while existing cross-domain (CDR) and multi-scenario (MSR) methods often require large amounts of resources and require strict input alignment, limiting their scalability. This paper proposes MTFM (Meituan Foundation Model for Recommendation), a Transformer-based framework designed to solve the above challenges. MTFM does not pre-align inputs, but converts cross-domain data into heterogeneous Tokens to capture multi-scenario knowledge in an alignment-free manner. In order to improve training efficiency, MTFM introduces a multi-scenario user-level sample aggregation mechanism, which significantly reduces the total number of instances and greatly improves training throughput; it also integrates Grouped-Query Attention and customized Hybrid Target Attention, which effectively reduces memory usage and computational complexity. In addition, the paper also implements a number of system-level optimizations, such as kernel fusion and elimination of CPU-GPU blocking, further improving training and inference throughput. Offline experiments in takeout and other scenarios have verified the effectiveness of MTFM, proving that significant performance improvements can be achieved by expanding model capacity and multi-scenario training data. Based on MTFM, the team built a unified base recommendation model that serves the above multiple main business scenarios, replaced their independent fine-scheduling models, and completed the full volume.
02 CDRRM: Contrast-Driven Rubric Generation for Reliable and Interpretable Reward Modeling
Paper download：PDF
Introduction to the paper: This paper proposes CDRRM, a comparison-driven scoring criterion generation and reward modeling framework, aiming to improve the reliability, interpretability and data efficiency of reward models in LLM alignment. Traditional reward models are “black box” and rely on expensive labeling; existing criterion methods are redundant and biased. CDRRM adopts a “comparison-aggregation” process: first compare good/poor answers to locate key differences, and then aggregate them into concise task-related criteria to guide the evaluation model. Experiments show that CDRRM reaches the state-of-the-art level on three benchmarks, alleviates biases such as talkativeness and location, and uses only 3,000 samples to allow the un-fine-tuned model to surpass the full fine-tuning baseline, which is both efficient and interpretable.
03 LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Services
Paper download：PDF
Introduction to the paper: This article aims at the research gaps in AI search in the field of local life services and constructs the LocalSearchBench evaluation benchmark. The benchmark covers 9 domestic cities and 6 major service categories, including 900 multi-hop question and answer tasks, and is equipped with the interactive environment LocalPlayground and the merchant search tool LocalRAG. After experimentally evaluating 16 mainstream large language reasoning models, it was found that the current models performed poorly on such tasks, and there were common problems such as insufficient information integrity and insufficient credibility. The research also analyzed typical defects such as model tool invocation and multi-hop reasoning, providing important support for model training and benchmark testing of agent search in local life service scenarios.
04 Deterministic-Allocation and Anonymous Joint Advertising in E-commerce Platforms
Paper download：PDF
Introduction to the paper: In order to solve the problem that the algorithm in the joint advertising auction scenario cannot satisfy anonymity and deterministic distribution at the same time, resulting in problems such as unfair distribution and incompatible incentives in practical applications, the JTransNet model was proposed. Anonymity requires that the auction results only depend on the bidding value and has nothing to do with the identity and order of the participants. Deterministic allocation ensures that the allocation result is unique under the same input. JTransNet implements an end-to-end data-driven automated AMD model auction algorithm by introducing an anonymity and deterministic allocation mechanism, combined with the differentiable NeuralSort sorting method. This algorithm well solves the problem of traffic allocation and deduction for multi-party investment in a joint auction scenario, and significantly improves platform advertising revenue in both offline and online experiments. JTransNet has been fully launched in Meituan’s core retail business scenarios, promoting the fairness and revenue of advertising business traffic sales, and at the same time providing an effective solution for the algorithm design of large-scale automated model auction mechanisms in the industry.
05 UME: A Unified Meta-Generalization Framework for Cross-Domain ETA
Paper download：PDF
Introduction to the paper: In real-time logistics scenarios, the estimated time of arrival (checkout page ETA) on the bill of lading page is crucial to improving user satisfaction, optimizing scheduling strategies, and controlling operating costs. On international instant delivery platforms such as Meituan Keeta, there is significant cross-domain heterogeneity, and multi-domain modeling has become a core requirement. However, existing methods face three major challenges: first, they cannot be generalized to completely unseen new market domains and cannot achieve zero-sample prediction in the cold start stage; second, cross-domain feature spaces are inconsistent, and new market domains have a structural lack of offline statistical features due to lack of historical data accumulation; third, mature domains and cold start domains often need to be modeled separately, hindering knowledge transfer and increasing maintenance costs. To this end, this article proposes UME (Unified Meta-generalization framework for ETA), a unified meta-generalization framework. UME designed a unified dual-branch network and a super-network-based meta-learning mechanism to dynamically modulate feature gating, expert attention and final prediction through domain-level knowledge and instance-level context to achieve cross-domain association modeling and intra-domain adaptation. At the same time, a knowledge distillation strategy is introduced to bridge the information gap caused by missing features. This method outperforms existing methods in both offline and online experiments.
06 Generative Large-Scale Pre-trained Models for Automated Ad Bidding Optimization
Paper download：PDF
Introduction to the paper: Modern automated bidding systems need to strike a balance between overall effectiveness, advertisers’ diverse goals and realistic constraints, reflecting the changing needs of the industry. In recent years, conditional generation models (such as Transformer and diffusion models) can directly generate bidding trajectories based on advertiser preferences, providing a promising alternative to traditional methods based on Markov decision processes. However, these generation methods also face challenges such as offline and online environment distribution shifts, limited action space exploration, and the need to meet constraints such as CPM and ROI. To this end, we propose GRAD, a scalable automated bidding base model. GRAD implements diversified bidding behavior exploration through the action hybrid expert module, and combines it with the causal Transformer for constrained optimization.
07 HMAF: A Hierarchical Multi-Slot GD-RTB Allocation Framework
Paper download：PDF
Introduction to the paper: In modern online advertising platforms, guaranteed delivery (GD) contracts and real-time bidding (RTB) auctions coexist and bid against each other. Existing methods either decouple the optimization of GD and RTB, or rely on heuristic priority rules. Therefore, they cannot effectively balance short-term revenue maximization and long-term contract delivery goals under complex multi-pit placement and exposure constraints. To address these issues, we proposed HMAF (Hierarchical Multi-Pit Allocation Framework), a unified framework designed to optimize exposure allocation in the GD-RTB advertising platform. HMAF takes the “planning-calibration-execution” paradigm as its core structure, combines offline constraint optimization with online decision-making, coordinates offline GD resource planning, dynamically calibrates the competition intensity between GD and RTB, and makes real-time list-level ranking decisions in a multi-pit environment.
08 MTGenRec: An Efficient Distributed Training System for Generative Recommendation Models in Meituan
Paper download：PDF
Introduction to the paper: Generative recommendations are increasingly widely used in search, recommendation, and advertising fields, and have achieved significant improvements in user experience and platform revenue. With the development trend of the generative recommendation model Scaling Dense, the industry is increasingly inclined to build a next-generation recommendation model training engine based on the PyTorch ecosystem and reuse the development dividends of LLM to the greatest extent. However, the PyTorch ecosystem still has a lot of room for improvement in supporting large-scale sparse Embedding training. Therefore, we proposed the MTGenRec training framework based on the PyTorch ecosystem to unify “sparse-dense” training capabilities and meet the training needs of industrial-level generative recommendation models. Specifically, for sparse IDs, we proposed to use dynamic Hashtable to replace the static table to solve the problem of dynamic ID going online and offline, making it easier for users to use. In order to improve training efficiency, we proposed automatic table merging, ID deduplication, variable length sequence load balancing and other technologies to optimize recommendation scenarios. In addition, we have also developed supporting technologies such as breakpoint resume training, mixed precision training, gradient accumulation, and operator fusion. A large number of experimental results show that compared to the TorchRec baseline, MTGenRec can achieve a training acceleration ratio of 1.6 times to 2.4 times while maintaining the same training accuracy. Expanding from 8 cards to 128 cards, MTGenRec has also achieved approximately linear expansion efficiency. Currently, MTGenRec has been used in multiple core scenarios within Meituan.
| Dianping Technology Department won the first and third place in the 2026 KDD Cup complex data analysis Data Agents international competition
KDD Cup is a globally recognized top competition in the field of data mining and knowledge discovery. In the DataAgents track of the 2026 season, the Meituan technical team took two months to break through the global participating teams and finally won the championship and third place.
The competition examines the model’s understanding and reasoning capabilities in real complex data scenarios – multi-modal input, unstructured documents, and interference information identification, each of which is the key to the implementation of the current Agent.
Although traditional Data+AI systems have made significant progress in specific tasks, the end-to-end analysis process still relies heavily on human expert orchestration, which has become a major bottleneck restricting the scalability and adaptability of data analysis. To this end, KDD Cup 2026 proposes to use “Data Agents” to break the situation – by integrating knowledge understanding, reasoning and planning capabilities, it can autonomously complete task disassembly and planning, tool selection and invocation, heterogeneous data reasoning and result synthesis. Complex data analysis is the core task, and the main challenges lie in the “heterogeneous gap” of real data and the nonlinear complexity of the inference link. The Data Agents intelligent agent receives a heterogeneous multi-modal data package (covering databases, PDF reports, JSON data, charts and even videos, etc.), and autonomously orchestrates a complex reasoning process including parallel branches, iterative loops and result aggregation for a high-level natural language question, and finally gives an accurate answer, aiming to promote the construction of a truly autonomous data analysis system.
During the competition, the team migrated the Agent Harness capabilities accumulated during the construction of the review “Ask Dian Zai” to the competition questions, and used their spare time to build a complete Agent runtime that supports independent exploration of multiple types of data files and the selection and execution of analysis tools such as SQL and Python. Through error feedback, timeout control, automatic retry and other mechanisms, the stability and fault tolerance of the agent in long-link tasks are improved. For heterogeneous data such as videos and unstructured documents, a multi-modal video understanding sub-agent and an unstructured document ETL sub-agent are also constructed to further enhance the data extraction, understanding and comprehensive analysis capabilities of the main agent. This award verifies the effectiveness of related technologies in complex heterogeneous data analysis agent scenarios, and also provides technical accumulation for the subsequent continuous improvement of the intelligence level of “Wandianzai”.
At present, the relevant code has been open sourced on GitHub. We will share more technical details through the technical blog in the future, so stay tuned!
