Autonomous Agents for Collaborative Task under Information Asymmetry

TL;DR

The paper tackles information asymmetry in multi-agent collaboration by introducing iAgents, a framework where agents retrieve and exchange human-derived information through InfoNav-guided planning and a mixed memory system (Clear + Fuzzy) to ensure accurate, comprehensive context. It defines InformativeBench to evaluate collaboration under asymmetry, featuring Needle-Oriented and Reasoning-Oriented tasks across large social networks and distributed reasoning challenges. Empirical results show GPT-4-based backends achieve meaningful performance with notable gains from recursive communication and memory integration, while smaller models struggle, underscoring both the promise and challenges of information-asymmetric MAS. The work highlights practical considerations for privacy, network modeling, and scalability, and points to future directions in edge deployment and human-in-the-loop verification for autonomous agent collaboration.)

Abstract

Large Language Model Multi-Agent Systems (LLM-MAS) have achieved great progress in solving complex tasks. It performs communication among agents within the system to collaboratively solve tasks, under the premise of shared information. However, when agents' collaborations are leveraged to perform multi-person tasks, a new challenge arises due to information asymmetry, since each agent can only access the information of its human user. Previous MAS struggle to complete tasks under this condition. To address this, we propose a new MAS paradigm termed iAgents, which denotes Informative Multi-Agent Systems. In iAgents, the human social network is mirrored in the agent network, where agents proactively exchange human information necessary for task resolution, thereby overcoming information asymmetry. iAgents employs a novel agent reasoning mechanism, InfoNav, to navigate agents' communication toward effective information exchange. Together with InfoNav, iAgents organizes human information in a mixed memory to provide agents with accurate and comprehensive information for exchange. Additionally, we introduce InformativeBench, the first benchmark tailored for evaluating LLM agents' task-solving ability under information asymmetry. Experimental results show that iAgents can collaborate within a social network of 140 individuals and 588 relationships, autonomously communicate over 30 turns, and retrieve information from nearly 70,000 messages to complete tasks within 3 minutes.

Figures (12)

• Figure 1: Comparison between previous MAS (left) and iAgents (right). The visibility range of information for each agent is highlighted with a colored background. On the left, all agents share all information (colored background of Virtual Company). On the right, each agent could only see information about its human user (separated colored backgrounds), and iAgents is designed to deal with such kind of information asymmetry.
• Figure 2: Overall architecture of iAgents. From left to right, 1) each individual in the social network is equipped with an agent, and 2) two human users invoke their agents to solve a task, each initially holding the information that is visible to its human user. Then 3) agents automatically raise communication and exchange necessary information on behalf of human users. Finally, 4) agents perform a consensus check on their planning completed by InfoNav to solve the task.
• Figure 3: A case of the task asking two agents to find the longest activity among all schedules. InfoNav navigates the communication by providing a plan to the agent. It first 1) asks the agent to make a plan on what information is needed, then 2) fills the placeholder in this plan during communication. Finally it 3) performs a consensus check on the completed plan to 4) get the answer.
• Figure 4: Two kinds of tasks in the InformativeBench. Each agent can only see the information (marked with different colors) of the human that it works on behalf of, which generates information asymmetry. Agents are 1) asked to find the needle information within the network or 2) reason to get an answer which is the output of an algorithm running on distributed information in the network.
• Figure 5: The figure depicts the distribution of different behaviors of agents in adjusting memory retrieval based on the progress of communication. Agents predominantly tend to maintain parameters unchanged, but when changes occur, they tend to increase parameters to gain more information.