#LI-Hybrid Team Overview

DiDi Global Inc. is the world’s leading mobility technology platform. It offers a wide range of app-based services across markets including Asia-Pacific, Latin America and Africa, including ride hailing, taxi hailing, chauffeur, hitch and other forms of shared mobility as well as auto solutions, food delivery, intra-city freight, and financial services.

Role Responsibilities

• Foundation Model Pre-training: Design and implement the pre-training pipeline for financial behavior sequence foundation models, including pre-training objective selection (CLM/MLM/Hybrid), Tokenization architecture experimentation (Flat/3D-Transformer/KVT), and scaling experiments.
• Multi-source Sequence Modeling: Build a unified sequence representation for behavioral data across multiple domains (payments, ride-hailing, food delivery, credit); design and validate the impact of data-source mixing ratios on model performance. Multi-entity and Multi-scale Fusion: Design an Account-Card dual-dimension sequence modeling scheme, along with a cross-temporal-scale fusion architecture bridging micro-level behaviors (millisecond-granularity event tracking) and macro-level behaviors (day/week-level transactions).
• Ablation Studies and Evaluation Framework: Build a systematic ablation experiment framework; design a freeze-backbone + linear head evaluation pipeline to drive architecture decisions.
• Production Deployment: Integrate pre-trained representations into downstream risk-control scenarios (stolen-card detection, credit scoring, etc.); design a Blending Module and complete SFT fine-tuning and online deployment.

Role Qualifications Must Have:

• Master's degree or above in Computer Science, Mathematics, Statistics, or a related field.
• 3+ years of deep learning algorithm R&D experience, with hands-on experience building a pre-trained model from scratch and completing the full training pipeline.
• Proficiency in Transformer architectures and variants (GPT/BERT/FT-Transformer/MoE), with practical sequence modeling experience.
• Familiar with at least one mainstream deep learning framework (PyTorch preferred); experience with distributed training (multi-GPU / multi-node).
• Solid experimental design skills: ability to independently conduct ablation studies and scaling-law experiments and draw reliable conclusions.
• Strong engineering implementation skills; able to iterate efficiently on model code and training pipelines.

Nice to Have:

• Modeling experience in financial risk control / anti-fraud / credit scoring.
• Publications on Foundation Models / Self-Supervised Learning (NeurIPS/ICML/ICLR/KDD/WWW, etc.).
• Familiarity with Contrastive Learning, ELECTRA, cross-modal fusion, and related techniques.
• Experience with time-series / event-sequence modeling (e.g., TimeMixer, TrajGPT).
• Experience with graph neural networks or graph-based anomaly detection.