EXHOLD: Experience-Aware Real-Time Hold Control for Large-Scale Ride-Hailing Matching at DiDi

2026-07-10 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

EXHOLD is a two-stage framework designed for real-time hold control in large-scale ride-hailing matching, specifically developed and deployed at DiDi in Brazil. This system addresses the limitations of traditional heuristic strategies, which often struggle with non-stationary traffic and multi-objective optimization for passenger-driver experience signals. EXHOLD's first stage involves learning a decision model that assigns each driver-order pair to discrete, interpretable experience tiers by optimizing a unified objective function that aggregates satisfaction signals across the entire matching funnel. The second stage then determines a monotone hold-time schedule through constrained optimization over empirical quantiles, explicitly enforcing service guardrails to prevent unnecessary delays for promising matches while maximizing overall experience improvement. Randomized A/B experiments in DiDi's production system demonstrated consistent gains, including increased trip completion, higher driver income, significantly reduced passenger cancellations, and improved funnel efficiency. Ablation studies confirmed the necessity of both stages for calibrated decision-making under spatiotemporal heterogeneity.

Key takeaway

For Machine Learning Engineers developing ride-hailing matching systems, EXHOLD demonstrates a robust approach to hold control. If you are struggling with brittle heuristic strategies under non-stationary traffic, consider implementing a two-stage framework that decouples experience assessment from hold-time execution. This method, proven at DiDi, can significantly increase trip completion, boost driver income, and reduce passenger cancellations, directly improving your platform's overall efficiency and user satisfaction.

Key insights

EXHOLD uses a two-stage, experience-aware framework to optimize ride-hailing hold control, improving efficiency and user satisfaction.

Principles

• Decouple assessment from execution.
• Optimize for unified experience signals.
• Enforce service guardrails explicitly.

Method

Stage I: Learn a decision model for driver-order pair experience tiers. Stage II: Solve for a monotone hold-time schedule via constrained optimization over empirical quantiles.

In practice

• Implement two-stage hold control.
• Aggregate multi-objective satisfaction signals.
• Use A/B tests for validation.

Topics

• Ride-hailing
• Hold Control
• DiDi
• Machine Learning
• A/B Testing
• Marketplace Efficiency

Best for: Research Scientist, AI Product Manager, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Reservation Based Smart Parking Management — Dynamic buffer and reputation systems enhance smart parking reliability by adapting to user behavior and incentivizing compliance.
• AI Customer Support at Scale: The Travel Industry’s $Billion Bet — The core challenge in AI customer support is managing the human-AI boundary, especially for complex adjudication cases.
• Regime-Calibrated Demand Priors for Ride-Hailing Fleet Dispatch and Repositioning — Matching current demand to historical regimes significantly reduces ride-hailing wait times without complex forecasting models.
• RAG Isn’t Enough — I Built the Missing Context Layer That Makes LLM Systems Work — Context engineering explicitly manages information flow into LLM context windows to prevent RAG system failures.
• Safety and alignment in an era of long-horizon models — Long-horizon models require iterative deployment, trajectory-level monitoring, and adaptive safeguards to manage emergent safety risks.
• Embark Now: User Demand Oriented Framework for Multi-day Urban Travel Itinerary Planning — Integrating LLMs for dynamic user preference capture with an enhanced GRASP algorithm optimizes complex multi-day travel planning.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.