L2M2: A Hierarchical Framework Integrating Large Language Model and Multi-agent Reinforcement Learning

1 Introduction

Multi-agent reinforcement learning (MARL) has emerged as a powerful framework for coordinating autonomous systems, demonstrating remarkable success in both real-world applications, such as urban traffic control [Zhang et al., 2019] and multi-robot navigation [Oroojlooy and Hajinezhad, 2023], and complex simulated environments such as StarCraft II [Vinyals et al., 2017; Samvelyan et al., 2019]. Recent studies have advanced MARL towards increasingly challenging problems, including environments with rich stochasticity [Ellis et al., 2023], human-agent collaboration [Li et al., 2023b;

∗Corresponding authors.

Liu et al., 2024a], and long-horizon sequential planning [Geng et al., 2024a]. These complex domains intensify challenges for MARL, such as sample efficiency, temporal credit assignment, exploration in large state-action spaces, and nonstationarity [Hernandez-Leal et al., 2019].

Hierarchical multi-agent architectures [Ahilan and Dayan, 2019; Geng et al., 2024b] attempt to address these challenges through multi-level control structures that decompose complex tasks into manageable subtasks. These methods typically employ a centralized high-level planner agent for subtask allocation and multiple low-level executor agents for subtask accomplishment. However, existing methods rely heavily on domain knowledge for subtask definition and assignment, resulting in task-specific policies with limited transferability across tasks. This dependency on domain knowledge and the costly retraining in new scenarios presents a critical bottleneck in scaling MARL to diverse real-world applications.

Recent research has explored integrating large language models (LLMs) into multi-agent systems [Li et al., 2023b], leveraging their pre-trained capabilities in contextual inference and natural language understanding. As existing approaches primarily focus on pure multi-LLM-agent systems, they generally face scalability challenges with large agent populations. To address these challenges, we propose L2M2 (Large Language Model and Multi-agent Reinforcement Learning), a hierarchical framework that integrates LLMs with multiple reinforcement learning agents. Following the LLMas-policy paradigm [Carta et al., 2023; Szot et al., 2024; Zhou et al., 2024], L2M2 introduces an LLM agent that leverages a large language model as its decision-making policy for task decomposition and allocation, while reinforcement learning agents execute the resulting subtasks through direct environmental interactions. L2M2 enables zero-shot planning [Dalal et al., 2023] to guide RL agents directly with pre-trained LLMs. This approach significantly reduces the computational overhead of developing policies for high-level controllers from scratch [Agarwal et al., 2023; Blumenkamp et al., 2024; Geng et al., 2024b] while enhancing the generalizability of L2M2 across different scenarios.

To validate our approach, we conduct comprehensive experiments across progressively challenging scenarios in VMAS [Bettini et al., 2022] and MOSMAC [Geng et al.,

2024a; Geng et al., 2025] environments, ranging from basic particle navigation to strategic navigation through obstaclerich terrains in StarCraft II maps. Our evaluation compares L2M2 against various baselines including classic nonhierarchical MARL algorithms and hierarchical frameworks, such as HiSOMA [Geng et al., 2024b] and MOSMAC's multiagent model with rule-based controller (RBC). The results demonstrate L2M2's significant advantages across multiple dimensions. In VMAS scenarios, L2M2 achieves comparable performance to baseline methods while using only 15-20% of training samples, demonstrating remarkable sample efficiency. On complex MOSMAC scenarios with the default pre-defined subgoals, L2M2 achieves a 98.75% success rate, outperforming both RBC (83.13%) and HiSOMA (94.38%). Most notably, while baseline methods struggle in the challenging MOSMAC scenarios without pre-defined subgoals, L2M2 maintains a 68.13% success rate through effective zero-shot planning [Dalal et al., 2023] and policy transfer using pretrained MARL models from the MOSMAC complex scenario with subgoals. Analysis using kernel density estimation reveals that L2M2's LLM agent consistently generates strategic navigation paths that avoid challenging areas, demonstrating effective high-level planning without domain expertise. These results highlight L2M2's potential as a generalizable framework for scaling up multi-agent systems towards long-horizon cooperative multi-agent navigation problems. The main contributions of this work are summarized as follows:

1. We propose L2M2, a novel hierarchical framework integrating LLM-based high-level planning with MARLbased execution, where an environment translator enables robust LLM-MARL coordination.
2. We introduce a zero-shot planning approach where pretrained LLMs directly guide MARL agents, supporting both end-to-end training and integration with existing MARL policies while reducing the cost of repetitively training high-level controllers for different tasks.
3. Experimental results on VMAS and MOSMAC demonstrate L2M2's effectiveness, achieving superior performance with only 15-20% of baseline training samples and maintaining high success rates (98.75% with pre-defined waypoints, 68.13% without) where baseline methods fail.

The rest of this paper is organized as follows. In Section 2, we discuss the relevant literature in the field. The L2M2 framework is detailed in Section 3. Section 4 presents the VMAS and MOSMAC benchmark environments. Section 5 discusses the experimental results. Section 6 concludes with a discussion of our findings and future work.

2 RelatedWork

2.1 LLM inMulti-agent Frameworks

Recent studies have demonstrated two primary approaches for integrating LLMs into multi-agent systems: LLM-as-policy and LLM-as-enhancement. The LLM-as-policy approach employs LLMs as decision-making policies for LLM agents , similar to neural networks in deep reinforcement learning agents. This approach leverages natural language as a communication protocol, enabling agents to incorporate perception, memory, and execution modules [Zhang et al., 2023]. In contrast, the LLM-as-enhancement approach uses LLMs to generate supplementary signals such as rewards that influence the agents' decision-making process. For a comprehensive review of this approach, we refer interested readers to a recent survey by Cao et al. [2024]. Multi-agent architectures on the LLM-as-policy approach have evolved along two main paths: the pure multiLLM-agent paradigm [Li et al., 2023a; Zhang et al., 2023; Liu et al., 2024b; Mandi et al., 2024; Huang et al., 2025] and the mixed multi-LLM/RL-agent paradigm [Li et al., 2023b; Liu et al., 2024a]. While pure systems focus on interactions between multiple LLM agents, mixed systems explore how reinforcement learning [Li et al., 2023b] or executable scripts [Yu et al., 2023] can complement LLM capabilities. This work introduces a novel perspective by investigating how language models can enhance the capabilities of low-level MARL agents through a mixed multi-LLM/RL framework, where the LLM agent follows the LLM-as-policy approach.

2.2 LLM-guided Reinforcement Learning

The mixed multi-LLM/RL-agent paradigm encompasses two complementary research directions. The first focuses on enhancing LLM agents through learned execution policies for plans and subtasks, while the second aims to improve the efficiency of reinforcement learning through LLM-guided policy learning. While substantial research has focused on the former, emerging studies investigate how LLMs can strengthen existing MARL algorithms. In the single-RL-agent domain, Dalal et al. [2023] showed the potential of using LLMs as high-level task planners to guide low-level policy training for robotic tasks. However, the extension to multi-agent scenarios remains largely unexplored. Specifically, L2M2 addresses the unexplored challenge of using LLMs to coordinate multiple reinforcement learning agents in complex navigation tasks, thereby filling a significant gap in the current literature on LLM-MARL integration.

2.3 MARL and Hierarchical Multi-agent Frameworks

MARL has made substantial progress in both on-policy and off-policy algorithms in recent years. Representative on-policy methods such as COMA [Foerster et al., 2018], MAA2C [Papoudakis et al., 2021], and MAPPO [Yu et al., 2022], along with off-policy methods including MADDPG [Lowe et al., 2017], VDN [Sunehag et al., 2018], QMIX [Rashid et al., 2018], QTRAN [Son et al., 2019], and MAVEN [Mahajan et al., 2019], have demonstrated strong performance on cooperative tasks. Building on these foundations, recent research has focused on scaling multi-agent systems to address long-horizon tasks. This progression, while promising, has intensified several key challenges, particularly temporal credit assignment [Jiang and Agarwal, 2018] and catastrophic forgetting [Elsayed and Mahmood, 2023]. Consequently, while existing MARL algorithms perform effectively in short-horizon scenarios such as SMAC [Samvelyan et al., 2019], they may face significant challenges when dealing with extended operation horizons that require sustained coordination and planning.

Hierarchical approaches have emerged as a promising direction for addressing these challenges. Methods such as Feudal Multiagent Hierarchies (FMH) [Ahilan and Dayan, 2019] and HAMMER [Gupta et al., 2023] demonstrate how task decomposition can enhance multi-agent capabilities. HiSOMA [Geng et al., 2024b] recently advances this approach by implementing a three-level control hierarchy that integrates multiple groups of RL agents trained through centralized MARL algorithms. This hierarchical framework has successfully tackled challenging long-horizon navigation problems that proved difficult for non-hierarchical approaches, highlighting the potential of hierarchical control structures to improve sample efficiency in policy learning for low-level RL agents.

3 The L2M2 Framework

We present L2M2, a hierarchical multi-agent framework that integrates large language models (LLMs) for high-level strategic planning with multi-agent reinforcement learning (MARL) for low-level execution. Our framework addresses the challenges of long-horizon task decomposition and coordination without requiring domain-specific heuristics or pre-defined subtasks. The framework consists of two primary components: (1) a high-level LLM agent that generates strategic plans and subtask allocations, and (2) multiple low-level RL agents that execute primitive actions to accomplish these subtasks. The system leverages pre-trained LLMs [Llama Team, 2024] for zero-shot planning while enabling both end-to-end training and direct integration with existing MARL policies.

3.1 Framework Overview

L2M2 implements a two-level hierarchy where a single LLM agent coordinates multiple RL agents through temporally extended [Vezhnevets et al., 2017] macro-actions. For the highlevel component, we employ the representative Llama 3.1-8B model [Llama Team, 2024], though our framework remains compatible with other LLMs. The low-level agents utilize state-of-the-art centralized training decentralized execution (CTDE) MARL algorithms, representatively QMIX [Rashid et al., 2018]. To facilitate efficient interaction between components, we implement a generalizable communication protocol comprising an environment translator ω and a subtask allocation mechanism. This modular design ensures flexibility, allowing each component to be implemented using any suitable model within its respective domain.

3.2 The Large Language Model Agent

The LLM agent serves as a high-level planner that performs zero-shot [Dalal et al., 2023] strategic planning by processing global environmental states sLLM and generating coordinated task allocations aLLM for subordinate RL agents. We adopt the LLM-as-policy [Carta et al., 2023; Szot et al., 2024; Zhou et al., 2024] approach, utilizing LLMs as policy networks πLLM that receive input states and produce output actions in textual forms. The LLM agent operates at a lower frequency than RL agents, which execute primitive actions at each timestep. Figure 1 illustrates this control process.

State Representation

Let ω : SEnv → P LLM be the environment translator function that maps environmental states into environmental prompts, where SEnv and P LLM are the environmental state space and the environmental prompt space respectively. At time t, the environmental state sEnv t ∈ SEnv is defined as:

where:

• ct is the current coordinates of n RL agents
• bt indicates obstacles in eight directions
• gt represents available movement subtasks
• dt denotes the current distances to the final target
• dt+1 indicates predicted distances after taking each potential subtask

The ω converts sEnv t into an environmental prompt:

where cLLM t , bLLM t , gLLM t , dLLM t , dLLM t+1 represent the natural language descriptions of their corresponding elements in sEnv t , and f specifies the format of LLM's response.

Action Space and Decision Making

The LLM agent generates temporally abstracted movement actions as subtasks for RL agents, illustrated in Figure 2.

The action aLLM t at time t is defined as:

where G is the discrete set of available subtasks for the next timestep t + 1 and n is the number of RL agents. To reduce computational overhead, L2M2 generates a joint plan for all RL agents in a single prompting round.

Feedback Mechanism

We implement a negative-feedback-only mechanism enabling the LLM agent to perform in-context learning. The verification process on aLLM is a two-step procedure that includes:

1. Hard verification: Checks output format compliance as specified by f in the prompt pLLM
2. Soft verification: Ensures selected actions in aLLM are within the available action space, reducing hallucinations

When verification fails, a self-correction procedure initiates another action selection round with error descriptions pErr incorporated into the environment prompt pLLM , allowing the LLM agent to reflect and correct its mistakes.

3.3 The Reinforcement Learning Agents

RL agents execute primitive actions that directly interact with the environment based on subtasks g assigned by the LLM agent. They maintain information-receiving channels in their observation space to process incoming communicative information from the high-level agent.

Observation and Action Spaces

At timestep t, RL agent i receives a partial observation:

where oe,i t represents general environment information, and og,i t encapsulates subtask-related information from the information-receiving channel. A translation module ωRL : Gi → Og,i maps subtask information regarding gi t+1 into the RL agents' observation space, ensuring effective communication of high-level strategic decisions.

The action space of RL agents remains identical to nonhierarchical MARL algorithms on the benchmark environments. For example, actions in the MOSMAC environment include no-op, move to four directions and stop.

Reward Structure

The reward function for RL agent i at timestep t combines environmental and subtask-related rewards:

where re,i t is the environmental reward and rg,i t is the subtaskrelated reward. This composite reward structure encourages RL agents to balance immediate environmental feedback with progress towards their assigned subtasks.

4 Benchmark Domains

To demonstrate the effectiveness of L2M2, this work compares L2M2 against baseline methods on two popular multi-agent environments, including VMAS [Bettini et al., 2022], a vectorized 2D physics environment that simulates the movements of multiple particles, and StarCraft II [Vinyals et al., 2017], a popular real-time strategy (RTS) game environment where agents need to learn to move and attack enemy units. For each environment, we adopt two tasks closely related to the multi-agent navigation problem. For the VMAS environment, we adopt the navigation and passage scenarios. For the StarCraft II environment, we adopt two long-horizon multi-agent navigation scenarios with complex terrain from MOSMAC [Geng et al., 2025], which provides a more realistic and challenging setting. In both environments, agents share the same goal and need to reach the goal collaboratively. The following subsections detail the specific characteristics of each environment.

(a) Complex w. subgoals (b) Complex w/o subgoals

Figure 4: Long-horizon multi-agent MOSMAC scenarios implemented in this study. In each scenario, four units spawn in the bottom-left region and navigate to the goal location (denoted by a building) in the upper-right region of a 128×128 map. For the scenario with subgoals, white spheres indicate pre-defined intermediary waypoints, with connecting lines representing traversable paths.

4.1 The VMAS Environment

VMAS (Vectorized Multi-Agent Simulator) [Bettini et al., 2022] provides a two-dimensional physics-based platform for simulating multi-agent interactions through particle-based dynamics. Among its scenarios, the navigation and passage scenarios (Figure 3) are particularly relevant for evaluating multi-agent navigation problems. The navigation scenario presents an obstacle-free environment where agents must cooperatively reach a shared goal position, while the passage scenario extends this by introducing a barrier that divides the map into two sections, with agents starting in the lower section and targeting a randomly placed goal in the upper section.

4.2 The MOSMAC Environment

The MOSMAC (Multi-Objective SMAC) benchmark [Geng et al., 2024a; Geng et al., 2025] (Figure 4) extends the StarCraft II Multi-Agent Challenge (SMAC) [Samvelyan et al., 2019] towards long-horizon and multi-objective problems. Our study focuses on MOSMAC scenarios featuring long-horizon multi-agent navigation: 4t_vs_0t_large_flat and 4t_vs_0t_large_complex (referred to as 'flat' and 'complex' hereafter), where four siege tank units navigate on a 128×128

map. Complex terrain scenarios incorporate challenging features such as high/low-ground elevation changes, ramps, cliffs, and dead-end corners that test agents' navigation capabilities in difficult areas. In addition to the original pre-defined scenario-specific subtasks, we also evaluate scenarios without these subgoals to create a more challenging environment that better reflects real-world applications.

5 Experiments and Results

We evaluate L2M2 against state-of-the-art baselines across progressively challenging scenarios in VMAS and MOSMAC environments, ranging from basic particle navigation to strategic navigation through obstacle-rich environments. The benchmark scenarios can be categorized based on whether agents should follow pre-defined subtasks, which influence the action space available to high-level controllers.

5.1 Baselines

For all baseline methods and L2M2, we employ QMIX [Rashid et al., 2018] as the base algorithm. QMIX is a state-of-the-art off-policy MARL algorithm that combines individual agent Q-values into a global Q-value through a mixing network. QMIX inspired several advanced algorithms including QTRAN [Son et al., 2019] and MAVEN [Mahajan et al., 2019], and has demonstrated superior performance on MOSMAC scenarios [Geng et al., 2024a]. While on-policy algorithms like MAA2C [Papoudakis et al., 2021] and MAPPO [Yu et al., 2022] could also be used for end-to-end training, their lower sample efficiency requires parallel training, which significantly increases computational costs for parallel LLM inference. Therefore, for on-policy algorithms, we focus on direct integration (DI) results where pre-trained MARL policies are combined with high-level controllers without additional training. All evaluations on L2M2 use Llama 3.1-8B [Llama Team, 2024] as the policy for its high-level controller. We set the temperature to 0 and limited the output to 150 tokens to ensure consistency.

We adopt HiSOMA [Geng et al., 2024b] as our primary hierarchical baseline. HiSOMA implements a multi-level control hierarchy where Fusion Architecture for Learning, Cognition, and Navigation (FALCON) [Tan, 2007] functions as a central higher-level controller that coordinates multiple groups of MARL agents. FALCON is based on the three-channel Fusion Adaptive Resonance Theory model, a type of self-organizing neural network (SONN) that learns state-action-reward associations for optimal control. Although HiSOMA was originally evaluated on MOSMAC with multiple groups of agents, we adapted it to our study by configuring a single FALCON controller to guide one group of RL agents through various navigation tasks, as a fair baseline for L2M2.

Scenarios without Pre-defined Subgoals

For scenarios without pre-defined subgoals, we evaluate L2M2 against both hierarchical and non-hierarchical methods through two approaches: end-to-end training of all methods, and direct integration of high-level controllers with MARL policies trained in a non-hierarchical manner. To ensure a fair comparison, all high-level controllers are configured with enhanced environmental interaction capabilities, with observation and action ranges set to eight times those of low-level RL agents (Figure 2). Following Geng et al. [2024b], we train HiSOMA's FALCON controller across four simulated scenarios that mirror the tasks for high-level controllers in VMAS and MOSMAC scenarios before integrating with MARL.

Scenarios with Subgoals

Scenarios with subgoals present a distinct evaluation setting, as they require agents to complete pre-defined subtasks in sequence. The MOSMAC benchmark provides a baseline method known as RBC-MARL, which employs a randomized path selector to guide units through strategic points toward their final goal (Figure 4b). This built-in method is essentially a hierarchical end-to-end training method where a path selector coordinates MARL agents through sequential subtask allocations. Given that pre-defined subtask selection is inherently a hierarchical control problem, we focus our evaluation on comparing L2M2 against other hierarchical approaches - specifically HiSOMA and the RBC-MARL baseline. We further demonstrate L2M2's effectiveness by showing it can directly utilize MARL policies trained with RBC without fine-tuning. This highlights its zero-shot adaptation capabilities.

5.2 Evaluation Protocols

We conduct comprehensive evaluations comparing L2M2 against established baselines while maintaining consistent hyperparameter settings across MARL algorithms. For nonhierarchical baselines and HiSOMA's end-to-end training, we allocate 1 million steps for VMAS navigation, 1.5 million steps for VMAS passage, and 20 million steps for MOSMAC complex terrain scenarios. Given L2M2's superior sample efficiency, we restrict its MARL training to fewer steps: 200k for VMAS navigation, 400k for VMAS passage, and 100k timesteps for MOSMAC scenarios, which reduces computational costs associated with LLM inference while maintaining performance. For end-to-end training evaluation, we assess all methods across 32 episodes at fixed intervals throughout the training process. For hierarchical methods using pre-trained MARL policies, we conduct five independent evaluations, each utilizing a MARL model trained with a different seed. Performance assessment relies on two standard MARL metrics: averaged win rates [Samvelyan et al., 2019] and averaged episode returns [Papoudakis et al., 2021], providing a comprehensive view of both task completion and reward optimization.

5.3 Results on the VMAS Scenarios

Figure 5 presents the results of L2M2 and the baseline methods on the VMAS navigation scenario. The non-hierarchical QMIX baseline achieves a 95.63% win rate within 1M steps. When integrated as the low-level controllers in hierarchical frameworks, it enables both HiSOMA and L2M2 to maintain comparable performance (96.25% and 95.00% win rates, respectively), demonstrating effective planning capabilities of both FALCON and Llama 3.1-8B controllers.

Notable differences emerge in end-to-end training. FALCON-guided training converges to an 84.38% win rate, suggesting a potential difficulty of policy learning under the commands of HiSOMA. In contrast, L2M2-guided training achieves a 94.79% win rate, suggesting more consistent and learnable guidance from the LLM controller. Most significantly, L2M2 demonstrates superior sample efficiency, converging to QMIX-level performance within 200k steps - only one-fifth of the samples required by non-hierarchical MARL.

These benefits become more pronounced in the more challenging VMAS passage scenario, where agents must learn sequential navigation through a passage to reach their goals (Figure 3b). As shown in Figure 6, L2M2 on average achieves an 8.75% win rate within 200k steps, while non-hierarchical MARL requires 1.5M steps to reach a lower 6.25% win rate. While HiSOMA encounters significant challenges in simto-real policy transfer, L2M2 achieves robust performance through its zero-shot planning [Dalal et al., 2023] capabilities, demonstrating its ability to effectively coordinate and guide multi-agent learning on complex tasks.

5.4 Results on the MOSMAC Scenarios

We first evaluate L2M2 on the default complex scenario with pre-defined subgoals, where a rule-based controller (RBC) selects from three pre-defined paths connecting start and goal locations. As presented in Table 1, end-to-end training with RBC-controlled QMIX agents achieves an average win rate of 83.13%. When integrating the trained QMIX policies with high-level controllers, both HiSOMA and L2M2 demonstrate significant improvements, achieving win rates of 94.38% and 98.75% respectively. These improvements can be attributed to enhanced planning capabilities that help agents avoid local optima and challenging terrain features that often trap RBCcontrolled agents. Notably, L2M2 exhibits enhanced planning capabilities compared to HiSOMA's pre-trained controller, demonstrating its efficient zero-shot planning ability.

We further evaluate L2M2 on a more challenging variant without pre-defined subgoals. In this scenario, nonhierarchical MARL algorithms struggle to learn effective navigation behaviour, resulting in poor performance of low-level controllers that preclude direct integration with high-level controllers. To address this limitation, we adopt a policy transfer approach, leveraging MARL models pre-trained in scenarios with subgoals. While both scenarios share identical terrains, they present distinct hidden environment transition functions and state distributions, creating a challenging zero-shot transfer setting. As presented in Table 2, L2M2 achieves a win rate of 68.13%, substantially outperforming HiSOMA, suggesting the efficiency of combining generalizable subtask-based MARL policies with LLMs in zero-shot settings.

5.5 Analysis on LLMAgent's Actions in L2M2

To further evaluate L2M2's capabilities, we analyzed the spatial distribution of the LLM agent's action selections using kernel-density estimation (KDE) with Gaussian kernels on the MOSMAC complex scenario without pre-defined subgoals, as illustrated in Figure 7. Notably, the LLM agent consistently generates navigation commands that avoid challenging terrain features such as cliffs, ramps, and elevation changes, which would otherwise trap the RL agents. Instead, it guides agents through the terrain's central region - a path that experienced human players would recognize as optimal for this navigation task. This spatial analysis demonstrates the LLM agent's ability to generate contextually appropriate subtask signals for the RL agents without requiring domain-specific training.

6 Conclusion and FutureWork

This paper introduces L2M2, a novel hierarchical multi-agent framework that integrates LLMs with MARL to address complex, long-horizon multi-agent tasks. Our approach demonstrates superior performance and sample efficiency, requiring only 15-20% of the training samples needed by baseline methods while achieving comparable results. The effectiveness of L2M2 stems from its zero-shot planning mechanism, which enables an LLM agent to generate contextually appropriate subtasks for MARL agents without repetitive retraining.

While demonstrating strong performance, L2M2 faces several limitations, including computational overhead from LLM and constraints on the LLM agent's discrete action space. Looking forward, several promising research directions emerge from this work. These include exploring LLMs' environmental perception for automatic target identification, incorporating feedback mechanisms for LLM fine-tuning, and scaling to larger multi-agent systems through more sophisticated coordination protocols. These advancements could significantly expand the capabilities of integrated LLM-MARL systems while maintaining the sample efficiency and generalization benefits demonstrated in this work.

Acknowledgments

This research was supported in part by Tsinghua University - Migu Xinkong Culture Technology (Xiamen) Co., Ltd. Joint Research Center for Intelligent Light Field and Interactive Technology, Beijing, under its Joint research and development project (R24119F0) - Phase 1, the National Research Foundation, Singapore, under its AI Singapore Programme (AISG Award No: AISG2-RP-2020-019), and the Lee Kong Chian Professorship awarded to Ah-Hwee Tan by Singapore Management University.

Contribution Statement

Minghong Geng: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Visualization, Writing - original draft, Writing - review & editing; Shubham Pateria: Software, Writing - review & editing; Budhitama Subagdja: Software, Writing - review & editing; Lin Li: Resources; Xin Zhao: Writing - review & editing, Funding Acquisition, Supervision; Ah-Hwee Tan: Conceptualization, Methodology, Resources, Writing - review & editing, Supervision, Project administration, Funding acquisition.