CROSS: A Mixture-of-Experts Reinforcement Learning Framework for Generalizable Large-Scale Traffic Signal Control

Xibei Chen; Yifeng Zhang; Yuxiang Xiao; Mingfeng Fan; Maonan Wang; Guillaume Sartoretti

cs.ROMar 26, 2026

CROSS: A Mixture-of-Experts Reinforcement Learning Framework for Generalizable Large-Scale Traffic Signal Control

Authors: Xibei Chen, Yifeng Zhang, Yuxiang Xiao, Mingfeng Fan, Maonan Wang, Guillaume Sartoretti

Organizations: Department of Mechanical Engineering, National University of Singapore, Singapore · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, Hong Kong

Abstract

Recent advances in robotics, automation, and artificial intelligence have enabled urban traffic systems to operate with increasing autonomy towards future smart cities, powered in part by the development of adaptive traffic signal control (ATSC), which dynamically optimizes signal phases to mitigate congestion and optimize traffic. However, achieving effective and generalizable large-scale ATSC remains a significant challenge due to the diverse intersection topologies and highly dynamic, complex traffic demand patterns across the network. Existing RL-based methods typically use a single shared policy for all scenarios, whose limited representational capacity makes it difficult to capture diverse traffic dynamics and generalize to unseen environments. To address these challenges, we propose CROSS, a novel Mixture-of-Experts (MoE)-based decentralized RL framework for generalizable ATSC. We first introduce a Predictive Contrastive Clustering (PCC) module that forecasts short-term state transitions to identify latent traffic patterns, followed by clustering and contrastive learning to enhance pattern-level representation. We further design a Scenario-Adaptive MoE module that augments a shared policy with multiple experts, thus enabling adaptive specialization and more flexible scenario-specific strategies. We conduct extensive experiments in the SUMO simulator on both synthetic and real-world traffic datasets. Compared with state-of-the-art baselines, CROSS achieves superior performance and generalization through improved representation of diverse traffic scenarios.

Figures & tables

Fig. 1: (a) Overview of a typical T-junction, which consists of six incoming lanes, six outgoing lanes, 12 traffic movements, and three traffic phases. (b) Traffic state and phase state vectors used in CROSS, with elements of both constructed based on ordered traffic movements.

Fig. 2: Architecture of the proposed CROSS framework, which includes a GFE module to extract rich traffic dynamics and form a unified representation for heterogeneous intersections, a PCC module to predict traffic dynamics and match them with learnable cluster centers for generalizable pattern representations, and a Scenario-Adaptive MoE module that uses these representations as routing context to compute weights and generate the traffic policy and value functions.

Traffic Dataset	Volume (veh)	Arrival Rate (veh/min)
Traffic Dataset	Volume (veh)	Mean	Std.	Max	Min
Grid $4\times 4$	1473.00	24.55	13.89	72.00	6.00
Arterial $4\times 4$	2484.00	41.40	24.69	88.00	10.00
Grid $5\times 5$	7296.00	121.60	121.94	752.00	32.00
$Jinan_{(1)}$	6295.00	104.92	19.79	136.00	50.00
$Jinan_{(2)}$	4365.00	72.75	15.15	101.00	43.00

TABLE I: Traffic demand specifications of the experimental datasets.

Method	Synthetic Dataset Evaluation					Real-World Dataset Evaluation (Zero-shot)
Method	Queue Length $\downarrow$ (veh)	Speed $\uparrow$ (m/s)	Trip Comp. Rate $\uparrow$ (veh/s)	Trip Time $\downarrow$ (s)	Trip Delay $\downarrow$ (s)	Queue Length $\downarrow$ (veh)	Speed $\uparrow$ (m/s)	Trip Comp. Rate $\uparrow$ (veh/s)	Trip Time $\downarrow$ (s)	Trip Delay $\downarrow$ (s)
Grid $4\times 4$ (Easy, Synthetic)						$Jinan_{(1)}$ (Hard, Real-World)
Fixed-Time	0.12(0.09)	7.88(1.28)	0.40(0.41)	194.63(76.87)	54.96(39.15)	2.02(0.93)	5.23(1.35)	1.36(0.65)	461.97(368.94)	170.23(249.98)
Max-Pressure	0.07(0.06)	8.82(0.66)	0.40(0.35)	174.41(63.07)	33.18(23.61)	0.33(0.14)	8.68(0.37)	1.60(0.62)	288.10(142.99)	25.95(28.56)
GESA	0.04(0.04)	10.08(0.73)	0.42(0.37)	154.03(53.46)	19.12(19.70)	0.94(0.36)	7.33(0.70)	1.56(0.60)	341.84(200.87)	76.53(97.96)
Unicorn	0.05(0.04)	9.55(0.59)	0.40(0.35)	160.38(53.93)	22.45(20.89)	3.81(1.98)	4.04(1.65)	1.29(0.51)	593.21(513.83)	313.28(389.58)

TABLE II: Performance comparison across synthetic training datasets and real-world evaluation scenarios.

\downarrow

denotes lower is better,

\uparrow

denotes higher is better. Best values are bold , second-best are underlined .

$Hangzhou_{(1)}$ (Medium, Real-World)
Method	Real-World Dataset Evaluation
Method	Queue Length $\downarrow$ (veh)	Speed $\uparrow$ (m/s)	Trip Comp. Rate $\uparrow$ (veh/s)	Trip Time $\downarrow$ (s)	Trip Delay $\downarrow$ (s)
Fixed-Time	0.95(0.47)	5.23(1.55)	0.65(0.38)	575.47(530.17)	226.95(409.00)
Max-Pressure	0.06(0.03)	9.47(0.28)	0.76(0.37)	332.22(169.81)	12.45(13.45)
GESA	0.25(0.10)	8.43(0.53)	0.75(0.37)	374.42(264.41)	57.89(192.77)
Unicorn	0.53(0.19)	6.90(0.69)	0.74(0.37)	458.48(298.37)	121.98(157.47)

TABLE III: Zero-shot Performance cross Hangzhou datasets.

Fig. 3: Average trip duration across three synthetic training datasets comparing CROSS with baseline methods (GESA, Unicorn) and ablations (CROSS w/o MoE, CROSS w/o PCC). Here

\downarrow

denotes lower is better.

Traffic Signal Control

Traffic signal control studies how to set intersection phases and green times to reduce congestion while balancing safety, emissions, and priority needs.

See all 39 papers

LATS: Large Language Model Assisted Teacher-Student Framework for Multi-Agent Reinforcement Learning in Traffic Signal Control
Mar 25, 2026Yifeng Zhang, Peizhuo Li, Tingguang Zhou +2Traffic Signal Control LLM-Guided RL
Department of Mechanical Engineering, National University of Singapore
A Robust and Efficient Multi-Agent Reinforcement Learning Framework for Traffic Signal Control
Mar 12, 2026Sheng-You Huang, Hsiao-Chuan Chang, Yen-Chi Chen +6Centralized Training with Decentralized Execution Reinforcement Learning
National Yang Ming Chiao Tung University · Academia Sinica · Kochi University of Technology +1
ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control
May 28, 2026Aoyu Pang, Maonan Wang, Yuejiao Xie +3RL Control Intelligent Transportation Systems
School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong · Shanghai AI Laboratory, Shanghai, China +1

Mar 25, 2026J/K move · Enter open · S save

LATS: Large Language Model Assisted Teacher-Student Framework for Multi-Agent Reinforcement Learning in Traffic Signal Control

Yifeng Zhang, Peizhuo Li, Tingguang Zhou, Mingfeng Fan+1

Department of Mechanical Engineering, National University of Singapore

Traffic Signal Control LLM-Guided RL Multi-Agent Reinforcement Learning

Open paper AskPDF arXiv

Adaptive Traffic Signal Control (ATSC) aims to optimize traffic flow and minimize delays by adjusting traffic lights in real time. Recent advances in Multi-agent Reinforcement Learning (MARL) have shown promise for ATSC, yet existing approaches still suffer from limited representational capacity, often leading to suboptimal performance and poor generalization in complex and dynamic traffic environments. On the other hand, Large Language Models (LLMs) excel at semantic representation, reasoning, and analysis, yet their propensity for hallucination and slow inference speeds often hinder their direct application to decision-making tasks. To address these challenges, we propose a novel learning paradigm named LATS that integrates LLMs and MARL, leveraging the former's strong prior knowledge and inductive abilities to enhance the latter's decision-making process. Specifically, we introduce a plug-and-play teacher-student learning module, where a trained embedding LLM serves as a teacher to generate rich semantic features that capture each intersection's topology structures and traffic dynamics. A much simpler (student) neural network then learns to emulate these features through knowledge distillation in the latent space, enabling the final model to operate independently from the LLM for downstream use in the RL decision-making process. This integration significantly enhances the overall model's representational capacity across diverse traffic scenarios, thus leading to more efficient and generalizable control strategies. Extensive experiments across diverse traffic datasets empirically demonstrate that our method enhances the representation learning capability of RL models, thereby leading to improved overall performance and generalization over both traditional RL and LLM-only approaches. [...]

Abstract

Traffic Signal Control

Explore similar work

LATS: Large Language Model Assisted Teacher-Student Framework for Multi-Agent Reinforcement Learning in Traffic Signal Control

A Robust and Efficient Multi-Agent Reinforcement Learning Framework for Traffic Signal Control

ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control