cs.AIAug 20, 2026

GenMatch: An End-to-End Generative Matching Framework for Micro-View Order-Dispatching in Ride-Hailing

Authors: Chuang Liu, Yuxueqing Zhang, Tengfei Lyu, Zirui Yuan, Weiqi Hu, Yanghan Cheng, Ming Wang, Li Ma, 

Organizations: Didi Chuxing Hangzhou, China · Didi Chuxing Beijing, China · The Hong Kong University of Science and Technology (Guangzhou) Guangzhou, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Micro-View Order-Dispatching assigns available drivers to passenger orders within each dispatch batch and is critical to the service quality and operational efficiency of ride-hailing platforms. Mainstream industrial solutions follow a multi-stage paradigm of model prediction, value calculation, and dispatch matching. Although dispatch quality is determined by the final batch-level assignment, these stages optimize different intermediate objectives. This cross-stage objective inconsistency means that improving a single stage does not necessarily improve the overall dispatch result. We therefore formulate Micro-View Order-Dispatching as a generative matching problem and propose GenMatch, an end-to-end Generative Matching framework and the first such framework deployed in a real-world production environment. Applying generative modeling to this problem introduces three challenges. First, each dispatch batch forms a dynamic sparse bipartite graph, requiring efficient structured batch-level encoding. Second, replacing the hand-crafted value function requires learning unified business utility from heterogeneous feedback. Third, directly generating an assignment requires tracking the evolving matching state because each selected order-driver pair changes the remaining feasible candidates. GenMatch addresses these challenges with a Context-Aware Bipartite Encoder, a Business-Aware Utility Learner, and a State-Aware Pointer Decoder. Extensive offline evaluations and online A/B tests in five cities across DiDi's international ride-hailing markets show consistent improvements over competitive baselines, confirming the effectiveness and practicality of GenMatch for industrial order-dispatching.

Figures & tables

Figure 1. Comparison of (a) the end-to-end generative paradigm and (b) the conventional multi-stage paradigm for Micro-View Order-Dispatching. (a) The proposed end-to-end generative dispatch framework, which models each dispatch batch as a whole and directly generates the dispatch result in an autoregressive manner. (b) The conventional multi-stage dispatch paradigm, which performs pair-level prediction, value calculation, and dispatch matching in separate stages.
Figure 2. GenMatch architecture: (a) Context-Aware Bipartite Encoder, (b) Business-Aware Utility Learner, and (c) State-Aware Pointer Decoder. The GenMatch architecture consists of a Context-Aware Bipartite Encoder, a Business-Aware Utility Learner, and a State-Aware Pointer Decoder. The decoder updates the selected and residual matching states after generating each OD pair.
VariantCity ICity IICity III
AR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrowAR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrowAR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrow
PDP KM0.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
PDP Greedy−- 1.86 ±\pm 0.05−- 1.69 ±\pm 0.09−- 0.24 ±\pm 0.02−- 0.14 ±\pm 0.01−- 0.97 ±\pm 0.06−- 1.26 ±\pm 0.04−- 0.25 ±\pm 0.01−- 1.16 ±\pm 0.06−- 1.92 ±\pm 0.15−- 2.35 ±\pm 0.18−- 0.31 ±\pm 0.02−- 0.19 ±\pm 0.02
PDP GS−- 0.42 ±\pm 0.02−- 0.30 ±\pm 0.03−- 0.06 ±\pm 0.01−- 0.04 ±\pm 0.00−- 0.38 ±\pm 0.02−- 0.54 ±\pm 0.02++ 0.10 ±\pm 0.01−- 0.47 ±\pm 0.04−- 0.79 ±\pm 0.06−- 0.76 ±\pm 0.07++ 0.14 ±\pm 0.01−- 0.27 ±\pm 0.02
D2SN−- 0.09 ±\pm 0.03++ 0.18 ±\pm 0.07−- 0.36 ±\pm 0.02−- 0.05 ±\pm 0.00−- 0.03 ±\pm 0.09−- 0.19 ±\pm 0.11−- 0.12 ±\pm 0.01−- 0.47 ±\pm 0.06−- 0.37 ±\pm 0.12−- 0.44 ±\pm 0.07−- 0.03 ±\pm 0.01−- 0.24 ±\pm 0.04
RLW−- 0.78 ±\pm 0.04−- 0.50 ±\pm 0.05++ 0.42 ±\pm 0.02−- 0.02 ±\pm 0.00−- 0.24 ±\pm 0.03−- 0.48 ±\pm 0.02++ 0.34 ±\pm 0.03−- 0.27 ±\pm 0.03−- 0.68 ±\pm 0.08−- 0.71 ±\pm 0.08++ 0.22 ±\pm 0.03−- 0.16 ±\pm 0.02
Table 1. Offline performance relative to PDP KM . Values are the mean ±\pm standard deviation of percentage changes over five runs; bold and underline denote the best and second-best results.
ModuleVariantAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrow
GenMatchFull0.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
EncoderA1−- 2.55 ±\pm 0.12−- 3.07 ±\pm 0.07++ 1.16 ±\pm 0.02−- 2.28 ±\pm 0.15
A2−- 2.43 ±\pm 0.06−- 2.69 ±\pm 0.03++ 1.14 ±\pm 0.01−- 2.15 ±\pm 0.06
A3−- 0.39 ±\pm 0.06−- 0.25 ±\pm 0.03++ 0.18 ±\pm 0.02−- 0.40 ±\pm 0.07
LearnerA4−- 2.76 ±\pm 0.16−- 3.15 ±\pm 0.08++ 1.17 ±\pm 0.03−- 2.32 ±\pm 0.18
A5−- 2.02 ±\pm 0.08−- 2.34 ±\pm 0.04++ 1.53 ±\pm 0.02−- 1.88 ±\pm 0.09
Table 2. Core City III ablations relative to GenMatch (Full), reported as the mean ±\pm standard deviation of percentage changes over five runs. Bold denotes the best result in each column.
VariantCity IIICity IVCity VOverall
AR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrowAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrowAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrowAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrow
GenMatchValue\mathbf{GenMatch}_{\mathrm{Value}}0.88%2.31% ∗−\mathbf{-} 7.08% ∗2.35% ∗0.59%3.06% ∗−\mathbf{-} 4.55% ∗2.43% ∗0.75% ∗1.67% ∗−\mathbf{-} 4.05% ∗1.04%0.77% ∗1.93% ∗−\mathbf{-} 4.85% ∗1.49% ∗
GenMatch3.18% ∗5.37% ∗−- 2.20% ∗4.89% ∗1.72% ∗4.51% ∗−- 1.68% ∗3.93% ∗2.01% ∗3.26% ∗−- 1.76% ∗2.16% ∗2.26% ∗3.86% ∗−\mathbf{-} 1.84% ∗2.97% ∗
Table 3. Online A/B test improvements over PDP KM (T −- C). Overall averages the three cities; ∗ indicates p<0.05p<0.05 .
MetricDelta (T −- C)
Dispatch Effectiveness Measures
Broadcast Count−- 0.17% ∗
Answer Count ( ↑\uparrow )2.16% ∗
Completion Count ( ↑\uparrow )3.84% ∗
Passenger Experience Measures
Passenger Bad Experience Ratio (PBE) ( ↓\downarrow )−- 15.17% ∗
Table 4. Online changes in dispatch effectiveness and experience relative to PDP KM (T −- C); ∗ indicates p<0.05p<0.05 .
Figure 3. Online gains of GenMatch across supply–demand periods (T −- C vs. PDP KM ). A grouped bar chart comparing low-, moderate-, and peak-demand periods. Completion-ratio gains increase from 3.24 percent to 3.96 percent and 4.12 percent. Passenger cancellation-after-answer ratio decreases by 1.76 percent, 7.71 percent, and 8.26 percent. Driver cancellation-after-answer ratio decreases by 3.48 percent, 4.31 percent, and 7.96 percent.
Figure 4. Production architecture of the Batch-Level Generative Dispatch Engine. The production architecture contains a control plane that collects requests, coordinates distributed shards, and assembles dispatch batches, and a compute plane that performs distributed feature extraction, candidate retrieval, and GenMatch inference.
SplitCityDaily Completed OrdersDaily Online DriversAvg. Trip Distance (m)
OfflineCity I3.04e+030.18e+033.47e+03
City II1.12e+040.55e+034.41e+03
Offline & OnlineCity III5.91e+030.43e+033.99e+03
OnlineCity IV1.73e+030.31e+034.29e+03
City V1.38e+040.68e+034.71e+03
Table 5. Scale statistics of the five experimental cities.
ModuleVariantCity ICity IICity III
AR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrowAR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrowAR (%) ↑\uparrowCR (%) ↑\uparrowAPT (%) ↓\downarrowGMV (%) ↑\uparrow
GenMatchFull0.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
EncoderA1−- 2.32 ±\pm 0.04−- 2.61 ±\pm 0.05++ 1.74 ±\pm 0.02−- 0.57 ±\pm 0.04−- 2.14 ±\pm 0.07−- 2.55 ±\pm 0.04++ 1.23 ±\pm 0.04−- 2.07 ±\pm 0.10−- 2.55 ±\pm 0.12−- 3.07 ±\pm 0.07++ 1.16 ±\pm 0.02−- 2.28 ±\pm 0.15
A2−- 2.25 ±\pm 0.02−- 2.30 ±\pm 0.02++ 1.61 ±\pm 0.01−- 0.52 ±\pm 0.02−- 1.95 ±\pm 0.03−- 2.36 ±\pm 0.02++ 1.06 ±\pm 0.02−- 1.96 ±\pm 0.05−- 2.43 ±\pm 0.06−- 2.69 ±\pm 0.03++ 1.14 ±\pm 0.01−- 2.15 ±\pm 0.06
A3−- 0.14 ±\pm 0.02−- 0.12 ±\pm 0.02++ 0.13 ±\pm 0.01−- 0.04 ±\pm 0.01−- 0.66 ±\pm 0.05−- 1.04 ±\pm 0.04++ 0.49 ±\pm 0.03−- 0.72 ±\pm 0.10−- 0.39 ±\pm 0.06−- 0.25 ±\pm 0.03++ 0.18 ±\pm 0.02−- 0.40 ±\pm 0.07
LearnerA4−- 2.47 ±\pm 0.05−- 2.69 ±\pm 0.06++ 2.03 ±\pm 0.01−- 0.59 ±\pm 0.05−- 2.20 ±\pm 0.09−- 2.97 ±\pm 0.06++ 1.28 ±\pm 0.05−- 2.30 ±\pm 0.15−- 2.76 ±\pm 0.16−- 3.15 ±\pm 0.08++ 1.17 ±\pm 0.03−- 2.32 ±\pm 0.18
Table 6. Complete ablation results relative to GenMatch (Full). Values are the mean ±\pm standard deviation of percentage changes over five runs.
VariantCity ICity IICity III
AUCDA\mathbf{AUC}_{\mathrm{DA}} (%) ↑\uparrowAUCPCAA\mathbf{AUC}_{\mathrm{PCAA}} (%) ↑\uparrowAUCDCAA\mathbf{AUC}_{\mathrm{DCAA}} (%) ↑\uparrowAUCDA\mathbf{AUC}_{\mathrm{DA}} (%) ↑\uparrowAUCPCAA\mathbf{AUC}_{\mathrm{PCAA}} (%) ↑\uparrowAUCDCAA\mathbf{AUC}_{\mathrm{DCAA}} (%) ↑\uparrowAUCDA\mathbf{AUC}_{\mathrm{DA}} (%) ↑\uparrowAUCPCAA\mathbf{AUC}_{\mathrm{PCAA}} (%) ↑\uparrowAUCDCAA\mathbf{AUC}_{\mathrm{DCAA}} (%) ↑\uparrow
PDP0.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
GenMatchMTL\mathbf{GenMatch}_{\mathrm{MTL}}++ 1.62 ±\pm 0.11++ 0.84 ±\pm 0.04++ 0.74 ±\pm 0.07++ 1.79 ±\pm 0.14++ 0.76 ±\pm 0.05++ 0.45 ±\pm 0.03++ 1.74 ±\pm 0.15++ 0.82 ±\pm 0.05++ 0.55 ±\pm 0.03
Table 7. Auxiliary prediction AUC relative to PDP, reported as the mean ±\pm standard deviation of percentage changes over five runs.
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5. Effect of model capacity relative to the Medium configuration. Error bars denote standard deviations over five runs. Four panels show changes in answer ratio, completion ratio, average pickup time, and gross merchandise volume relative to the Medium configuration across three cities. Six configurations vary encoder depth, decoder depth, and hidden dimension.
λmtl\lambda_{\mathrm{mtl}}City ICity IICity III
AR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrowAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrowAR ↑\uparrowCR ↑\uparrowAPT ↓\downarrowGMV ↑\uparrow
0.01−- 0.29 ±\pm 0.05−- 0.44 ±\pm 0.06++ 0.67 ±\pm 0.01−- 0.19 ±\pm 0.05−- 0.27 ±\pm 0.05−- 0.29 ±\pm 0.06++ 0.43 ±\pm 0.01−- 0.57 ±\pm 0.05−- 0.78 ±\pm 0.05−- 0.88 ±\pm 0.06++ 0.22 ±\pm 0.01−- 0.60 ±\pm 0.05
0.1−- 0.17 ±\pm 0.02−- 0.39 ±\pm 0.02++ 0.37 ±\pm 0.01−- 0.08 ±\pm 0.02−- 0.19 ±\pm 0.02−- 0.22 ±\pm 0.02++ 0.24 ±\pm 0.01−- 0.24 ±\pm 0.02−- 0.49 ±\pm 0.02−- 0.46 ±\pm 0.02++ 0.16 ±\pm 0.01−- 0.35 ±\pm 0.02
1−- 0.11 ±\pm 0.03−- 0.07 ±\pm 0.03++ 0.20 ±\pm 0.01−- 0.02 ±\pm 0.03−- 0.08 ±\pm 0.03−- 0.14 ±\pm 0.03++ 0.12 ±\pm 0.01−- 0.18 ±\pm 0.03−- 0.20 ±\pm 0.03−- 0.17 ±\pm 0.03++ 0.05 ±\pm 0.01−- 0.11 ±\pm 0.03
100.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.000.00 ±\pm 0.00
100++ 0.21 ±\pm 0.05−- 0.11 ±\pm 0.06++ 0.13 ±\pm 0.01++ 0.05 ±\pm 0.05−- 0.07 ±\pm 0.05−- 0.16 ±\pm 0.06−- 0.13 ±\pm 0.01−- 0.06 ±\pm 0.05++ 0.08 ±\pm 0.05−- 0.33 ±\pm 0.06−- 0.08 ±\pm 0.01−- 0.28 ±\pm 0.05
Appendix
Table 8. Sensitivity to λmtl\lambda_{\mathrm{mtl}} relative to the selected value 10. Values are the mean ±\pm standard deviation of percentage changes over five runs. Larger AR, CR, GMV, and AUC and smaller APT are preferred. Bold and underlined values denote the best and second-best results in each column, respectively.
ConfigurationSymbolValue
Encoder layersLencL_{\mathrm{enc}}2
Decoder layersLdecL_{\mathrm{dec}}2
Hidden dimensiondd128
Matching-attention heads–4
Competition-attention heads–4
Pointer headsPptrP_{\mathrm{ptr}}4
Appendix
Table 9. Model and training configurations of GenMatch.
More in this topic

Combinatorial Optimization

Combinatorial optimization finds the best feasible choices among discrete alternatives in routing, scheduling, graph problems, and resource allocation.

See all 384 papers

Explore similar work

CardsList

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

    Jul 10, 2026Xu Liu, Kai Wan, Zihao LuMulti-Objective Optimization

    Didichuxing Co. Ltd · Beijing, China

  2. Integrated Order Dispatching and Routing for Last-Mile Pickup via Deep Reinforcement Learning

    Jul 24, 2026Yida Xu, Zhaofang Mao, Yuheng Miao +2Vehicle RoutingSupply Chain Management

    College of Management and Economics, Tianjin University, Tianjin 300072, China · Laboratory of Computation and Analytics of Complex Management Systems (CACMS), Tianjin University, Tianjin 300072, China

  3. ProfiLLM: Utility-Aligned Agentic User Profiling for Industrial Ride-Hailing Dispatch

    Jun 17, 2026Tengfei Lyu, Zirui Yuan, Xu Liu +4User ModelingLLM Agent Training

    HKUST(GZ) · Didichuxing Co. Ltd