arXiv subjects

Zhexuan Wang

Publications and source records attributed to Zhexuan Wang.

5 recordsLinked to original sources

arXiv2026

MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

Large language model (LLM)-based Multi-agent systems (MAS) have shown promise in tackling complex collaborative tasks, where agents are typically orchestrated via role-specific prompts. While the quality of these prompts is pivotal, jointly optimizing them across interacting agents remains a non-trivial challenge, primarily due to the misalignment between local agent objectives and holistic system goals. To address this, we introduce MASPO, a novel framework designed to automatically and iteratively refine prompts across the entire system. A core innovation of MASPO is its joint evaluation mechanism, which assesses prompts not merely by their local validity, but by their capacity to facilitate downstream success for successor agents. This effectively bridges the gap between local interactions and global outcomes without relying on ground-truth labels. Furthermore, MASPO employs a data-driven evolutionary beam search to efficiently navigate the high-dimensional prompt space. Extensive empirical evaluations across 6 diverse tasks demonstrate that MASPO consistently outperforms state-of-the-art prompt optimization methods, achieving an average accuracy improvement of 2.9. We release our code at https://github.com/wangzx1219/MASPO.

cs.AI↗

arXiv2025

AgentInit: Initializing LLM-based Multi-Agent Systems via Diversity and Expertise Orchestration for Effective and Efficient Collaboration

Proper initialization is crucial for any system, particularly in multi-agent systems (MAS), where it plays a pivotal role in determining both the system's efficiency and effectiveness. However, existing MAS initialization methods do not fully account for the collaborative needs of the generated agents in subsequent stages. Inspired by the principles of effective team composition, we propose AgentInit, which aims to optimize the structure of agent teams. Specifically, in addition to multi-round interactions and reflections between agents during agent generation, AgentInit incorporates a Natural Language to Format mechanism to ensure consistency and standardization. Balanced team selection strategies using Pareto principles are subsequently applied to jointly consider agent team diversity and task relevance to promote effective and efficient collaboration and enhance overall system performance. Experiments show that AgentInit consistently outperforms state-of-the-art initialization methods and pre-defined strategies across various frameworks and tasks, achieving an overall performance improvement of up to 1.2 and 1.6, respectively, while also significantly reducing token consumption. Further analysis confirms its strong transferability to similar tasks and verifies the effectiveness of its key components, demonstrating its capability and adaptability as a reliable MAS initialization method. Source code and models are available at https://github.com/1737423697/AgentInit.

cs.AI↗

arXiv2025

AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration

Multi-agent systems (MAS) based on large language models (LLMs) have demonstrated significant potential in collaborative problem-solving. However, they still face substantial challenges of low communication efficiency and suboptimal task performance, making the careful design of the agents' communication topologies particularly important. Inspired by the management theory that roles in an efficient team are often dynamically adjusted, we propose AgentDropout, which identifies redundant agents and communication across different communication rounds by optimizing the adjacency matrices of the communication graphs and eliminates them to enhance both token efficiency and task performance. Compared to state-of-the-art methods, AgentDropout achieves an average reduction of 21.6% in prompt token consumption and 18.4% in completion token consumption, along with a performance improvement of 1.14 on the tasks. Furthermore, the extended experiments demonstrate that AgentDropout achieves notable domain transferability and structure robustness, revealing its reliability and effectiveness. We release our code at https://github.com/wangzx1219/AgentDropout.

cs.CL↗

arXiv2024

Low-Voltage Electron Emission by Graphene-hBN-graphene Heterostructure

Scanning Electron Microscopes (SEM) with low energy electron sources (accelerating voltage of less than 1000V) have important application requirements in many application scenarios. Tunneling junction can potentially achieve low-voltage and planar-type electron sources with good emission current density. However, further lower the extracting voltage while ensure the emission current density remains challenging. In this paper, we report a low-voltage planar-type electron source based on graphene-hBN-graphene heterostructures (GBGH) under a really low out-plane extracting voltage. The external electric field strength applied to the electron sources is only 4 times 10^4V/m and the accelerating voltage as low as 20V is realized. Steady electron emission of over 1nA and operating duration of several hours is observed from the GBGH with size of 59.29um^2 in our experiments, and thus the maximum emission current density reaches 7mA/cm^2. Great electrical contacts, extremely low thickness, and excellent layer properties of two-dimensional (2D) materials lead to easy-fabrication and miniature on-chip electron sources, which would significantly contribute to the development of next-generation free electron devices.

physics.app-ph↗

arXiv2022

pMPL: A Robust Multi-Party Learning Framework with a Privileged Party

In order to perform machine learning among multiple parties while protecting the privacy of raw data, privacy-preserving machine learning based on secure multi-party computation (MPL for short) has been a hot spot in recent. The configuration of MPL usually follows the peer-to-peer architecture, where each party has the same chance to reveal the output result. However, typical business scenarios often follow a hierarchical architecture where a powerful, usually privileged party, leads the tasks of machine learning. Only the privileged party can reveal the final model even if other assistant parties collude with each other. It is even required to avoid the abort of machine learning to ensure the scheduled deadlines and/or save used computing resources when part of assistant parties drop out. Motivated by the above scenarios, we propose pMPL, a robust MPL framework with a privileged part}. pMPL supports three-party training in the semi-honest setting. By setting alternate shares for the privileged party, pMPL is robust to tolerate one of the rest two parties dropping out during the training. With the above settings, we design a series of efficient protocols based on vector space secret sharing for pMPL to bridge the gap between vector space secret sharing and machine learning. Finally, the experimental results show that the performance of pMPL is promising when we compare it with the state-of-the-art MPL frameworks. Especially, in the LAN setting, pMPL is around $16\times$ and $5\times$ faster than TF-encrypted (with ABY3 as the back-end framework) for the linear regression, and logistic regression, respectively. Besides, the accuracy of trained models of linear regression, logistic regression, and BP neural networks can reach around 97%, 99%, and 96% on MNIST dataset respectively.

cs.CR↗