mia-code 0.2.0 → 0.3.0

package/contextual_research/PDE-generalization--caefee82-efb1-4dbb-8733-691b01581464--260130/sources/footnote_1_1_coordination_challenges.md

@@ -0,0 +1,766 @@
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+Shanshan Han 1 Qifan Zhang 1 Weizhao Jin 2 Zhaozhuo Xu 3
+
+6
+2
+0
+2
+
+n
+a
+J
+
+8
+2
+
+]
+
+A
+M
+
+.
+s
+c
+[
+
+3
+v
+8
+7
+5
+3
+0
+.
+2
+0
+4
+2
+:
+v
+i
+X
+r
+a
+
+Abstract
+
+This paper explores multi-agent systems and
+identify challenges that remain inadequately ad-
+dressed. By leveraging the diverse capabilities
+and roles of individual agents, multi-agent sys-
+tems can tackle complex tasks through agent col-
+laboration. We discuss optimizing task alloca-tion, fostering robust reasoning through iterative
+debates, managing complex and layered context
+information, and enhancing memory management
+to support the intricate interactions within multi-
+agent systems. We also explore potential appli-
+cations of multi-agent systems in blockchain sys-
+tems to shed light on their future development and
+application in real-world distributed systems.
+
+1. Introduction
+
+Multi-agent systems enhance the capabilities of single LLM
+agents by leveraging collaborations among agents and their
+specialized abilities (Talebirad & Nadiri, 2023; Zhang et al.,
+2023a; Park et al., 2023; Li et al., 2023; Jinxin et al., 2023).
+It utilizing collaboration and coordination among agents to
+execute tasks that are beyond the capability of any individual
+agent. In multi-agent systems, each agent is equipped with
+distinctive capabilities and roles, collaborating towards the
+fulfillment of some common objectives. Such collaboration,
+characterized by activities such as debate and reflection, has
+proven particularly effective for tasks requiring deep thought
+and innovation. Recent works include simulating interactive
+environments (Park et al., 2023; Jinxin et al., 2023), role-
+playing (Li et al., 2023), reasoning (Du et al., 2023; Liang
+et al., 2023), demonstrating the huge potential of multi-agent
+systems in handling complex real-world scenarios.
+
+While existing works have demonstrated the impressive ca-
+pabilities of multi-agent systems, the potential for advanced
+multi-agent systems far exceeds the progress made to date.
+A large number of existing works focus on devising planning
+
+1University of California, Irvine, CA, USA 2University of
+Southern California, Los Angeles, CA, USA 3Stevens Institute of
+Technology, Hoboken, NJ, USA. Correspondence to: Shanshan
+Han <shanshan.han@uci.edu>.
+
+strategies within a single agent by breaking down the tasks
+into smaller, more manageable tasks (Chen et al., 2022; Ziqi
+& Lu, 2023; Yao et al., 2023; Long, 2023; Besta et al., 2023;
+Wang et al., 2022b). Yet, multi-agent systems involve agents
+of various specializations and more complex interactions
+and layered context information, which poses challenges to
+the designing of the work flow as well as the whole system.
+Also, existing literature pays limited attention to memory
+storage, while memory plays a critical role in collaborations
+between agents. It enables agents to access to some common
+sense, aligning context with their tasks, and further, learn
+from past work flows and adapt their strategies accordingly.
+
+To date, multiple significant challenges that differentiate
+multi-agent systems and single-agent systems remain inade-
+quately addressed. We summarize them as follows.
+
+• Optimizing task allocation to leverage agents’ unique
+
+skills and specializations.
+
+• Fostering robust reasoning through iterative debates
+or discussions among a subset of agents to enhance
+intermediate results.
+
+• Managing complex and layered context information,
+such as context for overall tasks, single agents, and
+some common knowledge between agents, while en-
+suring alignment to the general objective.
+
+• Managing various types of memory that serve for dif-
+ferent objectives in coherent to the interactions in multi-
+agent systems
+
+This paper explores multi-agent systems, offering a sur-
+vey of the existing works while shedding light on the chal-lenges and open problems in it. We study major compo-nents in multi-agent systems, including planning and mem-ory storage, and address unique challenges posed by multi-agent systems, compared with single-agent systems. We
+also explore potential application of multi-agent systems
+in blockchain systems from two perspectives, including 1)
+utilizing multi-agent systems as tools, and 2) assigning an
+agent to each blockchain node to make it represent the user,
+such that the agent can can complete some tasks on behalf
+of the user in the blockchain network.
+
+1
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+2. Overview
+
+2.1. Structure of Multi-agent Systems
+
+The structure of multi-agent systems can be categorized into
+various types, based on the each agent’s functionality and
+their interactions.
+
+Equi-Level Structure. LLM agents in an equi-level system
+operate at the same hierarchical level, where each agent
+has its role and strategy, but neither holds a hierarchical
+advantage over the other, e.g., DMAS (Chen et al., 2023);
+see Figure 1(a). The agents in such systems can have same,
+neutral, or opposing objectives. Agents with same goals
+collaborate towards a common goal without a centralized
+leadership. The emphasis is on collective decision-making
+and shared responsibilities (Li et al., 2019). With opposing
+objectives, the agents negotiate or debate to convince the
+others or achieve some final solutions (Terekhov et al., 2023;
+Du et al., 2023; Liang et al., 2023; Chan et al., 2023).
+
+Hierarchical Structure. Hierarchical structures (Gronauer
+& Diepold, 2022; Ahilan & Dayan, 2019) typically con-sists of a leader and one or multiple followers; see Fig-ure 1(b). The leader’s role is to guide or plan, while the
+followers respond or execute based on the leader’s instruc-tions. Hierarchical structures are prevalent in scenarios
+where coordinated efforts directed by a central authority
+are essential. Multi-agent systems that explore Stackel-berg games (Von Stackelberg, 2010; Conitzer & Sandholm,
+2006) fall into this category (Harris et al., 2023). This type
+of game is distinguished by this leadership-followership dy-namic and the sequential nature of decision-making. Agents
+make decisions in a sequential order, where the leader player
+first generate an output (e.g., instructions) then the follower
+players take an action based on the leader’s instruc-tion.
+
+Nested Structure. Nested structures, or hybrid structures,
+constitute sub-structures of equi-level structures and/or hier-archical structures in a same multi-agent system (Chan et al.,
+2023); see Figure 1(c). The “big picture” of the system can
+be either equi-level or hierarchical, however, as some agents
+have to handle complex tasks, they break down the tasks
+into small ones and construct a sub-system, either equi-level
+or hierarchical, and “invite” several agents to help with
+those tasks. In such systems, the interplay between different
+levels of hierarchy and peer-to-peer interaction contributes
+to complexity. Also, the interaction among those different
+structures can lead to intricate dynamics, where strategies
+and responses become complicated due to the presence of
+various influencing factors, including external elements like
+context or environment.
+
+Dynamic Structure. Dynamic structures mean that the
+states of the multi-agent system, e.g., the role of agents,
+their relations, and the number of agents in the multi-agent
+system, may change (Talebirad & Nadiri, 2023) over time.
+
+As an example,
+(Talebirad & Nadiri, 2023) enables ad-dition and removal of agents to make the system to suit
+the tasks at hand. A multi-agent system may also be con-textually adaptive, with the interaction patterns inside the
+system being modified based on internal system states or
+external factors, such as contexts. Agents in such systems
+can dynamically reconfigure their roles and relationships in
+response to changing conditions.
+
+2.2. Overview of Challenges in Multi-Agent Systems
+
+This paper surveys various components of multi-agent sys-tems and discusses the challenges compared with single-agent systems. We discuss planning, memory management,
+as well as potential applications of multi-agent systems on
+distributed systems, e.g., blockchain systems.
+
+Planning. In a single-agent system, planning involves the
+LLM agent breaking down large tasks into a sequence of
+small, manageable tasks to achieve specific goals efficiently
+while enhancing interpretability, controllability, and flexi-bility (Li et al., 2024; Zhang et al., 2023b; Nye et al., 2021;
+Wei et al., 2022). The agent can also learn to call external
+APIs for extra information that is missing from the model
+weights (often hard to change after pre-training), or connect
+LLMs with websites, software, and tools (Patil et al., 2023;
+Zhou et al., 2023; Cai et al., 2023) to help reasoning and
+improve performance. While agents in a multi-agent sys-tem have same capabilities with single-agent systems, they
+encounter challenges inherited from the work flow in multi-agent systems. In §3, we discuss partitioning work flow and
+allocating the sub-tasks to agents; we name this process as
+“global planning”; see §3.1. We then discuss task decom-position in each single-agent. Different from planning in a
+single-agent systems, agents in multi-agent systems must
+deal with more sophisticated contexts to reach alignment in-side the multi-agent system, and further, achieve consistency
+towards the overall objective; see §3.2.
+
+Memory management. Memory management in single-agent systems include short-term memory during a conver-sation, long-term memory that store historical conversations,
+and, if any, external data storage that serves as a complemen-tary information source for inferences, e.g., RAG (Lewis
+et al., 2020). Memory management in multi-agent systems
+must handle complex context data and sophisticated interac-tion and history information, thus requires advanced design
+for memories. We classify the memories involved in multi-agent systems in §4.1 and then discuss potential challenges
+posed by the sophisticated structure of memory in §4.2.
+
+Application. We discuss applications of multi-agent sys-tems in blockchain, a distributed system that involves sophis-ticated design of layers and applications. Basically, multi-agent systems can serve as a tool due to its ability to handle
+sophisticated tasks in blockchain; see §5.1. Blockchain can
+
+2
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+Figure 1. Structures of multi-agent systems.
+
+also be integrated with multi-agent systems due to their dis-tributed nature, where an intelligent agent can be allocated
+to an blockchain node to perform sophisticated actions, such
+as negotiations, on behalf of the agent; see §5.2.
+
+3. Planning
+
+Planning in multi-agent systems involves understanding the
+overall tasks and design work flow among agents based on
+their roles and specializations, (i.e., global planning) and
+breaking down the tasks for each agent into small manage-able tasks (i.e., local planning). Such process must account
+for functionalities of the agents, dynamic interactions among
+the agents, as well as a more complex context compared with
+single-agent systems. This complexity introduces unique
+challenges and opportunities in the multi-agent systems.
+
+3.1. Global Planning
+
+Global planning refers to understanding the overall task
+and split the task into smaller ones and coordinate the sub-tasks to the agents. It requires careful consideration of task
+decomposition and agent coordination. Below we discuss
+the unique challenges in global planning in multi-agent
+systems.
+
+Designing effective work flow based on the agents’ spe-cializations. Partitioning responsibilities and designing
+effective work flows for agents is crucial for ensuring that
+the tasks for each agent are executable while meaningful
+and directly contributes to the overall objective in systems.
+The biggest challenge lies in the following perspectives: 1)
+the partition of work flow should maximize the utilization of
+each agent’s unique capabilities, i.e., each agent can handle
+a part of the task that matches its capabilities and expertise;
+2) each agent’s tasks must align with the overall goal; and
+3) the design must understand and consider the context for
+the overall tasks as well as each agent. This requires a deep
+understanding of the task at hand and the specific strengths
+and limitations of each agent in the system.
+
+Introducing loops for a subset of agents to enhance in-termediate results. Multi-agent systems can be integrated
+
+with loops inside one or multiple subsets of agents to im-prove the quality of the intermediate results, or, local op-timal answers. In such loops, agents debate or discuss to
+achieve an optimal results that are accepted by the agents
+in the loop. The iterative process can refine the interme-diate results, leading to a deeper exploration of the task.
+The agents in the loop can adjust their reasoning process
+and plans during the loop, thus have better capabilities in
+handling uncertainties of the task.
+
+Game Theory. Game theory provides a well-structured
+framework for understanding strategic interactions in multi-agent systems, particularly for systems that involve com-plex interactions among agents such as debates or discus-sions. A crucial concept in game theory is equilibrium, e.g.,
+Nash Equilibrium (Kreps, 1989) and Stackelberg Equilib-rium (Von Stackelberg, 2010; Conitzer & Sandholm, 2006),
+that describes a state where, given the strategies of others,
+no agent benefits from unilaterally changing their strategy.
+Game theory has been applied in multi-agent systems, espe-cially Stackelberg equilibrium (Gerstgrasser & Parkes, 2023;
+Harris et al., 2023), as the structure of Stackelberg equilib-rium contains is a leader agent and multiple follower agents,
+and such hierarchical architectures are wildely considered in
+multi-agent systems. (Gerstgrasser & Parkes, 2023) designs
+a general multi-agent framework to identify Stackelberg
+Equilibrium in Markov games, and (Harris et al., 2023)
+extend the Stackelberg model to allow agents to consider
+external context information, such as traffic and weather, etc.
+However, some problems are still challenging in multi-agent
+systems, such as defining an appropriate payoff structure
+for both the collective strategy and individual agents based
+on the context of the overall tasks, and efficiently achiev-ing equilibrium states. These unresolved issues highlight
+the ongoing need for refinement in the application of game
+theory to complex multi-agent scenarios.
+
+3.2. Single-Agent Task Decomposition
+
+Task decomposition in a single agent involves generating
+a series of intermediate reasoning steps to complete a task
+or arrive at an answer. This process can be represented as
+
+3
+
+7a) Equi-level structureb) Hierarchical structurec) Nested structureLeaderFlowersLeaderFlowersLLM Multi-Agent Systems: Challenges and Open Problems
+
+transforming direct input-output (⟨input → output⟩) map-pings into the ⟨input → rational → output⟩ mappings (Wei
+et al., 2022; Zhang et al., 2023b). Task composition can be
+of different formats, as follows.
+
+i) Chain of Thoughts (CoT) (Wei et al., 2022) that transforms
+big tasks into step-by-step manageable tasks to represent
+interpretation of the agents’ reasoning (or thinking) process.
+
+ii) Multiple CoTs (Wang et al., 2022a) that explores multiple
+independent CoT reasoning paths and return the one with
+the best output.
+
+iii) Program-of-Thoughts (PoT) (Chen et al., 2022) that
+uses language models to generate text and programming
+language statements, and finally an answer.
+
+iv) Table-of-Thoughts (Tab-CoT) (Ziqi & Lu, 2023) that
+utilize a tabular-format for reasoning, enabling the com-plex reasoning process to be explicitly modelled in a highly
+structured manner.
+
+v) Tree-of-Thoughts (ToT) (Yao et al., 2023; Long, 2023)
+that extends CoT by formulating a tree structure to explore
+It enables
+multiple reasoning possibilities at each step.
+generating new thoughts based on a given arbitrary thought
+and possibly backtracking from it.
+
+vi) Graph-of-Thoughts-Rationale (GoT-Rationale) (Besta
+et al., 2023) that explores an arbitrary graph to enable ag-gregating arbitrary thoughts into a new one and enhancing
+the thoughts using loops.
+
+vii) Rationale-Augmented Ensembles (Wang et al., 2022b)
+that automatically aggregate across diverse rationales to
+overcome the brittleness of performance to sub-optimal
+rationales.
+
+In multi-agent systems, task decomposition for a single
+agent becomes more intricate. Each agent must understand
+layered and sophisticated context, including 1) the overall
+tasks, 2) the specific context of the agent’s individual tasks,
+and 3) the contextual information provided by other agents
+in the multi-agent system. Moreover, the agents must align
+these complex, multi-dimensional contexts into their decom-posed tasks to ensure coherent and effective functioning
+within the overall task. We summarize the challenges for
+single agent planning as follows.
+
+Aligning Overall Context. Alignment of goals among
+different agents is crucial in multi-agent systems. Each LLM
+agent must have a clear understanding of its role and how it
+fits into the overall task, such that the agents can perform
+their functions effectively. Beyond individual roles, agents
+need to recognize how their tasks fit into the bigger picture,
+such that their outputs can harmonize with the outputs of
+other agents, and, further, ensuring all efforts are directed
+towards the common goal.
+
+Aligning Context Between Agents. Agents in multi-agent systems process tasks collectively, and each agent
+must understand and integrate the contextual information
+provided by other agents within the system to ensure that
+the information provided by other agents is fully utilized.
+
+Aligning Context for Decomposed Tasks. When tasks of
+each agents are broken down into smaller, more manageable
+sub-tasks, aligning the complex context in multi-agent sys-tems becomes challenging. Each agent’s decomposed task
+must fit their individual tasks and the overall goal while inte-grating with contexts of other agents. Agents must adapt and
+update their understanding of the task in response to context
+provided by other agents, and further, plan the decomposed
+tasks accordingly.
+
+Consistency in Objectives. In multi-agent systems, con-sistency in objectives is maintained across various levels,
+i.e., from overall goals down to individual agent tasks and
+their decomposed tasks. Each agent must understand and
+effectively utilize the layered contexts while ensuring its
+task and the decomposed sub-tasks to remain aligned with
+the overall goals. (Harris et al., 2023) extends the Stackel-berg model (Von Stackelberg, 2010; Conitzer & Sandholm,
+2006) to enable agents to incorporate external context in-formation, such as context (or insights) provided by other
+agents. However, aligning the complex context with the
+decomposed tasks during reasoning remains unresolved.
+
+4. Agent Memory and Information Retrieval
+
+The memory in single-LLM agent systems refers to the
+agent’s ability to record, manage, and utilize data, such as
+past historical queries and some external data sources, to
+help inference and enhance decision-making and reason-ing (Yao et al., 2023; Park et al., 2023; Li & Qiu, 2023;
+Wang et al., 2023; Guo et al., 2023). While the memory
+in a single-LLM agent system primarily focuses on inter-nal data management and utilization, a multi-agent system
+requires agents to work collaboratively to complete some
+tasks, necessitating the individual memory capabilities of
+each agent as well as a sophisticated mechanism for sharing,
+integrating, and managing information across agents, thus
+poses challenges to memory and information retrieval.
+
+4.1. Classifications of Memory in Multi-agent Systems
+
+Based on the work flow of a multi-agent system, we catego-rize memory in multi-agent system as follows.
+
+• Short-term memory: This is the immediate, transient
+memory used by a Large Language Model (LLM) dur-ing a conversation or interaction, e.g., working memory
+in (Jinxin et al., 2023). It is ephemeral, existing only
+for the duration of the ongoing interaction and does
+not persist once the conversation ends.
+
+4
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+• Long-term Memory: This type of memory stores his-torical queries and responses, essentially chat histories
+from earlier sessions, to support inferences for future
+interactions. Typically, this memory is stored in exter-nal data storage, such as a vector database, to facilitate
+recall of past interactions.
+
+• External data storage: This is an emerging area in
+LLM research where models are integrated with ex-ternal data storage like vector databases, such that the
+agents can access additional knowledge from these
+databases, enhancing their ability to ground and enrich
+their responses (Lewis et al., 2020). This allows the
+LLM to produce responses that are more informative,
+accurate, and highly relevant to the specific context of
+the query.
+
+• Episodic Memory: This type of memory encompasses
+a collection of interactions within multi-agent systems.
+It plays a crucial role when agents are confronted with
+new tasks or queries. By referencing past interactions
+that have contextual similarities to the current query,
+agents can significantly enhance the relevance and ac-curacy of their responses. Episodic Memory allows for
+a more informed approach to reasoning and problem-solving, enabling a more adaptive and intelligent re-sponse mechanism, thus serves as a valuable asset in
+the multi-agent system,
+
+• Consensus Memory: In a multi-agent system where
+agents work on a task collaboratively, consensus mem-ory acts as a unified source of shared information, such
+as common sense, some domain-specific knowledge,
+etc, e.g., skill library in (Jinxin et al., 2023). Agents
+utilize consensus memory to align their understand-ing and strategies with the tasks, thus enhancing an
+effective and cohesive collaboration among agents.
+
+While both single-agent and multi-agent systems handle
+short-term memory and long-term memory, multi-agent sys-tems introduce additional complexities due to the need for
+inter-agent communication, information sharing, and adap-tive memory management.
+
+4.2. Challenges in Multi-agent Memory Management
+
+Managing memory in multi-agent systems is fraught with
+challenges and open problems, especially in the realms of
+safety, security, and privacy. We outline these as follows:
+
+Hierarchical Memory Storage: In a multi-agent system,
+different agents often have varied functionalities and access
+needs. Some agents may have to query their sensitive data,
+but they don’t want such data to be accessed by other parties.
+While ensuring the consensus memory to be accessible to
+all clients, implementing robust access control mechanisms
+is crucial to ensure sensitive information of an agent is not
+accessible to all agents. Additionally, as the agents in a sys-tem collaborative on one task, and their functionalities share
+same contexts, their external data storage and memories
+may overlap. If the data and functionalities of these agents
+are not sensitive, adopting an unified data storage can effec-tively manage redundancy among the data, and furthermore,
+ensure consistency across the multi-agent system, leading
+to more efficient and precise maintenance of memory.
+
+Maintenance of Consensus Memory: As consensus mem-ory is obtained by all agents when collaborating on a task,
+ensuring the integrity of shared knowledge is critical to en-sure the correct execution of the tasks in the multi-agent
+systems. Any tampering or unauthorized modification of
+consensus memory can lead to systemic failures of the ex-ecution. Thus, a rigorous access control is important to
+mitigate risks of data breaches.
+
+Communication and information exchange: Ensuring
+effective communication and information exchange between
+agents is essential in multi-agent systems. Each agent may
+hold critical pieces of information, and seamless integration
+of these is vital for the overall system performance.
+
+Management of Episodic Memory. Leveraging past inter-actions within the multi-agent system to enhance responses
+to new queries is challenging in multi-agent systems. De-termining how to effectively recall and utilize contextually
+relevant past interactions among agents for current problem-solving scenarios is important.
+
+These challenges underscore the need for continuous re-search and development in the field of multi-agent systems,
+focusing on creating robust, secure, and efficient memory
+management methodologies.
+
+5. Applications in Blockchain
+
+Multi-agent systems offer significant advantages to
+blockchain systems by augmenting their capabilities and
+efficiency. Essentially, these multi-agent systems serve as
+sophisticated tools for various tasks on blockchain and Web3
+systems. Also, blockchain nodes can be viewed as agents
+with specific roles and capabilities (Ankile et al., 2023).
+Given that both Blockchain systems and multi-agent sys-tems are inherently distributed, the blockchain networks
+can be integrated with multi-agent systems seamlessly. By
+assigning a dedicated agent to each blockchain node, it’s
+possible to enhance data analyzing and processing while
+bolstering security and privacy in the chain.
+
+5.1. Multi-Agent Systems As a Tool
+
+To cast a brick to attract jade, we give some potential direc-tions that multi-agents systems can act as tools to benefit
+blockchain systems.
+
+Smart Contract Analysis. Smart contracts are programs
+
+5
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+stored on a blockchain that run when predetermined con-ditions are met. Multi-agents work together to analyze
+and audit smart contracts. The agents can have different
+specializations, such as identifying security vulnerabilities,
+legal compliance, and optimizing contract efficiency. Their
+collaborative analysis can provide a more comprehensive
+review than a single agent could achieve alone.
+
+Consensus Mechanism Enhancement. Consensus mech-anisms like Proof of Work (PoW) (Gervais et al., 2016) or
+Proof of Stake (PoS) (Saleh, 2021) are critical for validating
+transactions and maintaining network integrity. Multi-agent
+systems can collaborate to monitor network activities, an-alyze transaction patterns, and identify potential security
+threats. By working together, these agents can propose
+enhancements to the consensus mechanism, making the
+blockchain more secure and efficient.
+
+Fraud Detection. Fraud detection is one of the most im-portant task in financial monitoring. As an example, (Ankile
+et al., 2023) studies fraud detection through the perspective
+of an external observer who detects price manipulation by
+analyzing the transaction sequences or the price movements
+of a specific asset. Multi-agent systems can benefit fraud de-tection in blockchain as well. Agents can be deployed with
+different roles, such as monitoring transactions for fraud-ulent activities and analyzing user behaviors. Each agent
+could also focus on different behavior patterns to improve
+the accuracy and efficiency of the fraud detection process.
+
+5.2. Blockchain Nodes as Agents
+
+(Ankile et al., 2023) identifies blockchain nodes as agents,
+and studies fraud detection in the chain from the perspective
+an external observer. However, as powerful LLM agents
+with analyzing and reasoning capabilities, there are much
+that the agents can do, especially when combined with game
+theory and enable the agents to negotiate and debate. Below
+we provide some perspectives.
+
+Smart Contract Management and Optimization. Smart
+contracts are programs that execute the terms of a contract
+between a buyer and a seller in a blockchain system. The
+codes are fixed, and are self-executed when predetermined
+conditions are met. Multi-agent systems can automate and
+optimize the execution of smart contracts with more flexible
+terms and even dynamic external information from users.
+Agents can negotiate contract terms on behalf of their users,
+manage contract execution, and even optimize gas fees (in
+the context of Ethereum (Wood et al., 2014). The agents can
+analyze context information , such as past actions and pre-defined criteria, and utilize the information with flexibility.
+Such negotiations can also utilize game theory, such as
+Stackelberg Equilibrium (Von Stackelberg, 2010; Conitzer
+& Sandholm, 2006) when there is a leader negotiator and
+Nash Equilibrium (Kreps, 1989) when no leader exists.
+
+6. Conclusion
+
+The exploration of multi-agent systems in this paper under-scores their significant potential in advancing the capabil-ities of LLM agents beyond the confines of single-agent
+paradigms. By leveraging the specialized abilities and col-laborative dynamics among agents, multi-agent systems can
+tackle complex tasks with enhanced efficiency and innova-tion. Our study has illuminated challenges that need to be
+addressed to harness the power of multi-agent systems bet-ter, including optimizing task planning, managing complex
+context information, and improving memory management.
+Furthermore, the potential applications of multi-agent sys-tems in blockchain technologies reveal new avenues for
+development, which suggests a promising future for these
+systems in distributed computing environments.
+
+References
+
+Ahilan, S. and Dayan, P. Feudal multi-agent hierarchies
+for cooperative reinforcement learning. arXiv preprint
+arXiv:1901.08492, 2019.
+
+Ankile, L., Ferreira, M. X., and Parkes, D. I see you! robust
+In Multi-Agent
+
+measurement of adversarial behavior.
+Security Workshop@ NeurIPS’23, 2023.
+
+Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Gi-aninazzi, L., Gajda, J., Lehmann, T., Podstawski, M.,
+Niewiadomski, H., Nyczyk, P., et al. Graph of thoughts:
+Solving elaborate problems with large language models.
+arXiv preprint arXiv:2308.09687, 2023.
+
+Cai, T., Wang, X., Ma, T., Chen, X., and Zhou, D.
+Large language models as tool makers. arXiv preprint
+arXiv:2305.17126, 2023.
+
+Chan, C.-M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S.,
+Fu, J., and Liu, Z. Chateval: Towards better llm-based
+evaluators through multi-agent debate. arXiv preprint
+arXiv:2308.07201, 2023.
+
+Chen, W., Ma, X., Wang, X., and Cohen, W. W. Program
+of thoughts prompting: Disentangling computation from
+reasoning for numerical reasoning tasks. arXiv preprint
+arXiv:2211.12588, 2022.
+
+Chen, Y., Arkin, J., Zhang, Y., Roy, N., and Fan, C. Scalable
+multi-robot collaboration with large language models:
+Centralized or decentralized systems? arXiv preprint
+arXiv:2309.15943, 2023.
+
+Conitzer, V. and Sandholm, T. Computing the optimal
+strategy to commit to. In Proceedings of the 7th ACM
+conference on Electronic commerce, pp. 82–90, 2006.
+
+6
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., and Mor-Improving factuality and reasoning in lan-datch, I.
+guage models through multiagent debate. arXiv preprint
+arXiv:2305.14325, 2023.
+
+Li, Y., Wen, H., Wang, W., Li, X., Yuan, Y., Liu, G., Liu,
+J., Xu, W., Wang, X., Sun, Y., et al. Personal llm agents:
+Insights and survey about the capability, efficiency and
+security. arXiv preprint arXiv:2401.05459, 2024.
+
+Gerstgrasser, M. and Parkes, D. C. Oracles & followers:
+Stackelberg equilibria in deep multi-agent reinforcement
+learning. In International Conference on Machine Learn-ing, pp. 11213–11236. PMLR, 2023.
+
+Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang,
+R., Yang, Y., Tu, Z., and Shi, S. Encouraging divergent
+thinking in large language models through multi-agent
+debate. arXiv preprint arXiv:2305.19118, 2023.
+
+Gervais, A., Karame, G. O., W¨ust, K., Glykantzis, V., Ritz-dorf, H., and Capkun, S. On the security and performance
+of proof of work blockchains. In Proceedings of the 2016
+ACM SIGSAC conference on computer and communica-tions security, pp. 3–16, 2016.
+
+Gronauer, S. and Diepold, K. Multi-agent deep reinforce-ment learning: a survey. Artificial Intelligence Review,
+pp. 1–49, 2022.
+
+Guo, Z., Cheng, S., Wang, Y., Li, P., and Liu, Y. Prompt-guided retrieval augmentation for non-knowledge-intensive tasks. arXiv preprint arXiv:2305.17653, 2023.
+
+Harris, K., Wu, S., and Balcan, M. F. Stackelberg games
+In Multi-Agent Security Work-with side information.
+shop@ NeurIPS’23, 2023.
+
+Jinxin, S., Jiabao, Z., Yilei, W., Xingjiao, W., Jiawen, L.,
+and Liang, H. Cgmi: Configurable general multi-agent
+interaction framework. arXiv preprint arXiv:2308.12503,
+2023.
+
+Kreps, D. M. Nash equilibrium. In Game Theory, pp. 167–177. Springer, 1989.
+
+Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V.,
+Goyal, N., K¨uttler, H., Lewis, M., Yih, W.-t., Rockt¨aschel,
+T., et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Pro-cessing Systems, 33:9459–9474, 2020.
+
+Li, G., Hammoud, H. A. A. K., Itani, H., Khizbullin, D., and
+Ghanem, B. Camel: Communicative agents for” mind”
+exploration of large scale language model society. arXiv
+preprint arXiv:2303.17760, 2023.
+
+Li, X. and Qiu, X. Mot: Memory-of-thought enables chatgpt
+to self-improve. In Proceedings of the 2023 Conference
+on Empirical Methods in Natural Language Processing,
+pp. 6354–6374, 2023.
+
+Li, X., Sun, M., and Li, P. Multi-agent discussion mech-anism for natural language generation. In Proceedings
+of the AAAI Conference on Artificial Intelligence, vol-ume 33, pp. 6096–6103, 2019.
+
+Long, J. Large language model guided tree-of-thought.
+
+arXiv preprint arXiv:2305.08291, 2023.
+
+Nye, M., Andreassen, A. J., Gur-Ari, G., Michalewski, H.,
+Austin, J., Bieber, D., Dohan, D., Lewkowycz, A., Bosma,
+M., Luan, D., et al. Show your work: Scratchpads for
+intermediate computation with language models. arXiv
+preprint arXiv:2112.00114, 2021.
+
+Park, J. S., O’Brien, J., Cai, C. J., Morris, M. R., Liang,
+P., and Bernstein, M. S. Generative agents: Interactive
+simulacra of human behavior. In Proceedings of the 36th
+Annual ACM Symposium on User Interface Software and
+Technology, pp. 1–22, 2023.
+
+Patil, S. G., Zhang, T., Wang, X., and Gonzalez, J. E. Gorilla:
+Large language model connected with massive apis. arXiv
+preprint arXiv:2305.15334, 2023.
+
+Saleh, F. Blockchain without waste: Proof-of-stake. The
+Review of financial studies, 34(3):1156–1190, 2021.
+
+Talebirad, Y. and Nadiri, A. Multi-agent collaboration:
+Harnessing the power of intelligent llm agents. arXiv
+preprint arXiv:2306.03314, 2023.
+
+Terekhov, M., Graux, R., Neville, E., Rosset, D., and Kolly,
+G. Second-order jailbreaks: Generative agents success-fully manipulate through an intermediary. In Multi-Agent
+Security Workshop@ NeurIPS’23, 2023.
+
+Von Stackelberg, H. Market structure and equilibrium.
+
+Springer Science & Business Media, 2010.
+
+Wang, W., Dong, L., Cheng, H., Liu, X., Yan, X., Gao, J.,
+and Wei, F. Augmenting language models with long-term
+memory. arXiv preprint arXiv:2306.07174, 2023.
+
+Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang,
+S., Chowdhery, A., and Zhou, D. Self-consistency im-proves chain of thought reasoning in language models.
+arXiv preprint arXiv:2203.11171, 2022a.
+
+Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., and
+Zhou, D. Rationale-augmented ensembles in language
+models. arXiv preprint arXiv:2207.00747, 2022b.
+
+7
+
+LLM Multi-Agent Systems: Challenges and Open Problems
+
+Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F.,
+Chi, E., Le, Q. V., Zhou, D., et al. Chain-of-thought
+prompting elicits reasoning in large language models.
+Advances in Neural Information Processing Systems, 35:
+24824–24837, 2022.
+
+Wood, G. et al. Ethereum: A secure decentralised gener-alised transaction ledger. Ethereum project yellow paper,
+151(2014):1–32, 2014.
+
+Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y.,
+and Narasimhan, K. Tree of thoughts: Deliberate prob-lem solving with large language models. arXiv preprint
+arXiv:2305.10601, 2023.
+
+Zhang, J., Xu, X., and Deng, S. Exploring collaboration
+mechanisms for llm agents: A social psychology view,
+2023a.
+
+Zhang, Z., Yao, Y., Zhang, A., Tang, X., Ma, X., He, Z.,
+Wang, Y., Gerstein, M., Wang, R., Liu, G., et al. Igniting
+language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents. arXiv preprint
+arXiv:2311.11797, 2023b.
+
+Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A.,
+Cheng, X., Bisk, Y., Fried, D., Alon, U., et al. Webarena:
+A realistic web environment for building autonomous
+agents. arXiv preprint arXiv:2307.13854, 2023.
+
+Ziqi, J. and Lu, W. Tab-CoT: Zero-shot tabular chain
+of thought.
+In Rogers, A., Boyd-Graber, J., and
+Okazaki, N. (eds.), Findings of the Association for Com-putational Linguistics: ACL 2023, pp. 10259–10277,
+Toronto, Canada, July 2023. Association for Computa-tional Linguistics. doi: 10.18653/v1/2023.findings-acl.
+651. URL https://aclanthology.org/2023.
+findings-acl.651.
+
+8
+