PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors

Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, Jie Zhou

arXiv:2308.10848v3cs.CL

TL;DR

Complex real-world tasks often require cooperation to improve effectiveness. AgentVerse models human group problem-solving and dynamically adjusts agent composition as work progresses. It outperforms individual agents across understanding, reasoning, coding, and tool-utilization tasks.

  • Problem

    Complex real-world tasks often require cooperation, motivating investigation into how autonomous agents can collaborate effectively.

    Source: §1 · p. 1

  • Method

    AgentVerse models human group problem-solving through expert recruitment, collaborative decision-making, action execution, and evaluation, dynamically adjusting group composition based on progress.

    Source: §1 · p. 2

  • Results

    AgentVerse outperforms individual agents across general understanding, reasoning, coding, and tool-utilization tasks.

    Source: §1 · p. 2, §6 · p. 10

  • Takeaways & Limitations

    Emergent volunteer behavior improves team efficiency, while conformity helps agents realign with the common goal following criticism from peers.

    Source: §1 · p. 2

    Hazardous emergent behaviors could pose risks in real-world deployment, particularly when humans participate, and strategies to prevent them remain a future research need.

    Source: §4.3 · p. 9

Abstract

from arXiv · show

Autonomous agents empowered by Large Language Models (LLMs) have undergone significant improvements, enabling them to generalize across a broad spectrum of tasks. However, in real-world scenarios, cooperation among individuals is often required to enhance the efficiency and effectiveness of task accomplishment. Hence, inspired by human group dynamics, we propose a multi-agent framework \framework that can collaboratively and dynamically adjust its composition as a greater-than-the-sum-of-its-parts system. Our experiments demonstrate that \framework framework can effectively deploy multi-agent groups that outperform a single agent. Furthermore, we delve into the emergence of social behaviors among individual agents within a group during collaborative task accomplishment. In view of these behaviors, we discuss some possible strategies to leverage positive ones and mitigate negative ones for improving the collaborative potential of multi-agent groups. Our codes for \framework will soon be released at \url{https://github.com/OpenBMB/AgentVerse}.

1 INTRODUCTION

AGENTVERSE models human-team problem-solving through dynamically composed agent groups and an iterative collaboration process. The introduction reports improvements across diverse capabilities and identifies both beneficial and harmful emergent social behaviors.

1 INTRODUCTION

  • The framework targets uncertain generalizability beyond narrow tasks and limited adaptability when agent roles and capabilities remain fixed.Complex real-world tasks motivate cooperation to improve effectiveness.Source: §1 · pp. 1–2
  • AGENTVERSE dynamically adjusts group composition through four stages: expert recruitment, collaborative decision-making, action execution, and evaluation.Recruited agents jointly devise strategies and execute actions in their environment. Evaluation compares the current state with desired outcomes and feeds unsatisfactory results into another refinement iteration.Source: §1 · p. 2
  • AGENTVERSE improves agents’ understanding, reasoning, coding, and tool-use capabilities, while demonstrating potential in embodied AI.The introduction supports these claims through experiments and case studies across these domains.Source: §1 · p. 2
  • Multi-agent collaboration produces volunteer, conformity, and destructive behaviors, particularly during tool use and Minecraft gameplay.Volunteering improves team efficiency through peer assistance; conformity aligns deviating behavior with shared goals following criticism. Destructive behaviors can produce undesired and detrimental outcomes.Source: §1 · p. 2

2 AGENTVERSE FRAMEWORK

AGENTVERSE organizes multi-agent problem-solving into four stages: Expert Recruitment, Collaborative Decision-Making, Action Execution, and Evaluation. When the goal remains unmet, evaluation feedback informs adjustments to the group’s composition for the next round.

2 AGENTVERSE FRAMEWORK

  • AGENTVERSE models its four-stage problem-solving process as a Markov decision process.The model comprises agent and environment states, solutions and actions, a transition function, a reward function, and goals.Source: §2 · p. 3

2.1 EXPERT RECRUITMENT

  • A recruiter agent dynamically generates goal-specific expert descriptions, forming a group whose composition changes with evaluation feedback rather than relying on predefined descriptions.Automating recruitment is intended to make agent configuration more scalable.Source: §2.1 · p. 3

2.2 COLLABORATIVE DECISION-MAKING

  • Collaborative decision-making uses either horizontal aggregation of agents’ refined decisions or vertical refinement by a solver responding to reviewers.Horizontal aggregation can use summarization or ensemble methods and suits consulting and tool use. Vertical refinement stops at reviewer consensus or an iteration limit and suits mathematics and software development.Source: §2.2 · p. 3

2.3 ACTION EXECUTION

  • During action execution, agents carry out collectively decided actions, transitioning the environment to a new state.Depending on the implementation, some agents may perform no execution.Source: §2.3 · p. 3

2.4 EVALUATION

  • Evaluation compares the current state with the goal and returns verbal feedback that guides recruitment adjustments when the goal remains unmet.Humans or an agent can define the feedback mechanism, which identifies shortcomings and suggests improvements; recruitment considers both this feedback and the goal.Source: §2.4 · p. 4

3 EXPERIMENTS

Zero-shot experiments show that AGENTVERSE improves general understanding, reasoning, coding, and multi-tool task completion over standalone agents. Collaboration also enriches outputs, although incorrect peer feedback can undermine GPT-3.5-Turbo’s reasoning.

3 EXPERIMENTS

  • Zero-shot experiments use GPT-3.5-Turbo-0613 and GPT-4-0613, comparing CoT, single-agent AGENTVERSE Solo, and collaborative AGENTVERSE Group configurations.Compared with CoT, Solo adds expert recruitment, action execution, and evaluation; Group additionally uses multiple agents during decision-making.Source: §3 · p. 4

3.1 GENERAL UNDERSTANDING AND REASONING CAPABILITIES

  • AGENTVERSE’s Solo and Group configurations consistently outperform standalone CoT on general understanding and reasoning benchmarks with either evaluated LLM.GPT-3.5-Turbo’s logical-reasoning results are omitted because it struggled with logic-grid puzzles.Source: §3.1 · p. 4
  • Roughly 10% of MGSM errors arise when GPT-3.5-Turbo agents abandon correct answers after incorrect peer feedback, a phenomenon absent in GPT-4 agents.GPT-3.5-Turbo’s Group configuration underperforms Solo on two of three tasks, motivating greater robustness to incorrect information during collaboration.Source: §3.1 · pp. 4–5
  • In hydrogen-storage consulting, GPT-4 expert collaboration produces deeper advice initially and broader, more detailed recommendations by the second round.Group lacks a clear benchmark advantage over Solo for either LLM; the authors attribute this partly to evaluation metrics’ limited scope.Source: §3.1 · p. 5

3.2 CODING CAPABILITIES

  • 83.5 to 89.0 is the reported GPT-4 improvement in HumanEval pass@1, with performance increasing from CoT through Solo to Group.GPT-3.5-Turbo also benefits from collaboration in coding; the authors hypothesize that extensive code pretraining increases resilience to erroneous feedback.Source: §3.2 · pp. 5–6, Table 2 · p. 6
  • The Group-built Python calculator improves interface usability and exception handling over Solo through feedback from designers, evaluators, and testers.Both versions achieve core functionality, but Group adds color distinctions and keyboard input. The authors also report efficiency, robustness, and security improvements beyond pass@1.Source: §3.2 · p. 6

3.3 TOOL UTILIZATION CAPABILITIES

  • 9 of 10 multi-tool tasks are completed by AGENTVERSE, versus 3 by standalone ReAct (Yao et al., 2023b).Each task requires at least two tools. In six of its seven failures, ReAct misses task criteria and exits prematurely.Source: §3.3 · p. 6
  • In the 24-point-game case, task division and evaluation feedback recover an omitted similar-games subtask that standalone ReAct leaves unanswered.The group ultimately supplies game rules, solving code with test cases, and a similar-game summary.Source: §3.3 · p. 7

4 EMERGENT BEHAVIORS WITHIN A MULTI-AGENT GROUP

AgentVerse agents exhibit volunteer, conformity, and destructive behaviors during collaborative Minecraft tasks. Volunteering accelerates shared goals and conformity restores task alignment, while efficiency-seeking destructive actions raise safety concerns.

4 EMERGENT BEHAVIORS WITHIN A MULTI-AGENT GROUP

  • Collaborative Minecraft crafting reveals volunteer, conformity, and destructive behaviors, which also appear in earlier experiments such as tool utilization.Minecraft tasks require agents to plan, coordinate, and adjust to evolving situations while collaboratively crafting items.Source: §4 · p. 7

4.1 VOLUNTEER BEHAVIORS

  • Agents volunteer spare time, share materials, and assist unfinished teammates, accelerating achievement of their shared goal.Agents propose parallel gathering instead of waiting and transfer materials to peers possessing crafting tools. After completing their assignments, agents offer to help teammates gather missing resources.Source: §4.1 · pp. 8–9

4.2 CONFORMITY BEHAVIOR

  • Peer criticism redirects a distracted agent toward the shared task, illustrating conformity that maintains alignment with group goals.After Charlie crafts unrelated items instead of gathering leather, Alice and Bob criticize his actions; he acknowledges the mistake and refocuses.Source: §4.2 · p. 9

4.3 DESTRUCTIVE BEHAVIOR

  • Agents sometimes harm other agents or destroy an entire village library to obtain materials more efficiently, raising safety concerns.Preventing hazardous behavior is identified as a critical research direction, particularly for real-world collaboration involving humans.Source: §4.3 · p. 9

5 RELATED WORK

AgentVerse builds on LLM-based autonomous agents and cooperative multi-agent systems. This section reviews mechanisms that augment individual agents, prior findings on collaboration, and AgentVerse’s ability to adjust group composition dynamically.

Autonomous Agents

  • External mechanisms address missing human-like cognitive capabilities in autonomous agents through reflection, task decomposition, and tool utilization or creation.Examples include reflection (Yao et al., 2023b; Shinn et al., 2023) and task decomposition (Wei et al., 2022b; Yao et al., 2023a). Tool utilization/creation mechanisms are explored in (Schick et al., 2023b; Qin et al., 2023a;b; Qian et al., 2023b).Source: §5 · p. 9
  • LLMs provide understanding, reasoning, and generation capabilities that enable autonomous agents to handle increasingly complex scenarios.The section identifies LLM advances (Touvron et al., 2023a; OpenAI, 2023a) and their use as agent backbones (Richards & et al., 2023; Nakajima, 2023; Reworkd, 2023; Liu et al., 2023).Source: §5 · p. 9

Multi-agent System

  • AgentVerse uses group cooperation for more intricate scenarios, dynamically adjusting its composition according to the current state to support decision-making and execution.The framework builds on prior findings about autonomous collaboration among agents.Source: §5 · p. 9
  • Prior studies investigate autonomous agent cooperation, reporting emergent social behaviors and improved reasoning performance.Li et al. (2023) explore cooperation within agent groups, while Park et al. (2023) report autonomously emerging social behaviors. Du et al. (2023); Wang et al. (2023b); Zhang et al. (2023a); Qian et al. (2023a); Chan et al. (2023) leverage cooperation to improve reasoning performance.Source: §5 · p. 9
  • Multi-agent systems pursue efficiency and effectiveness through cooperation, distributing decisions and actions across agents working toward a common goal.The section contrasts earlier multi-agent joint training with recent efforts toward autonomous collaboration.Source: §5 · p. 9

6 CONCLUSION

AGENTVERSE emulates human group problem-solving and outperforms individual agents across general understanding, reasoning, coding, and tool utilization. Minecraft investigations reveal both positive and negative emergent social behaviors among agents.

6 CONCLUSION

  • AGENTVERSE outperforms individual agents across general understanding, reasoning, coding, and tool utilization.With appropriate tools, it also achieves strong results on intricate user queries.Source: §6 · p. 10
  • AGENTVERSE is a general multi-agent framework designed to emulate human group problem-solving processes.The authors envision further adaptation and refinement for a wider range of tasks and contexts.Source: §6 · p. 10
  • Minecraft investigations identify both positive and negative emergent social behaviors among agents.The authors argue that understanding multi-agent interactions will become increasingly important as artificial general intelligence advances.Source: §6 · p. 10

A CONFIGURATIONS OF THE EXPERIMENTS

The experiments assess understanding, reasoning, coding, and tool use with task-specific metrics, agent counts, and collaboration structures. Execution feedback supports an evaluator that accepts correct decisions or sends feedback back to expert recruitment.

Datasets and Evaluation Metrics

Expert Recruitment

  • Four agents handle dialogue, code completion, and constrained generation; mathematical reasoning uses two, and tool use uses two or three depending on the task.Agent counts are predefined rather than automated. Mathematical reasoning uses fewer reviewers because additional reviewers were observed to increase erroneous critiques that led solvers to incorrect solutions.Source: A CONFIGURATIONS OF THE EXPERIMENTS · p. 16

Collaborative Decision-Making

  • Coding, understanding, and reasoning use a vertical structure with one answering solver; tool use employs a horizontal structure for discussing individual subtasks.The vertical structure suits tasks requiring a single response, whereas tool-using agents must clarify their respective subtasks during discussion.Source: A CONFIGURATIONS OF THE EXPERIMENTS · p. 16

Action Execution

  • Execution supplies unit-test results, missing concepts, and tool-task status reports to evaluation.For GPT-4 Humaneval, an additional agent generates zero-shot unit tests; test results and missing Commongen-Challenge concepts become environment states for evaluation. Tool agents use ReAct for at most 10 iterations, then must conclude with a task status of “pending” or “finished” for evaluator assessment.Source: A CONFIGURATIONS OF THE EXPERIMENTS · p. 16

Evaluation

  • An evaluator agent checks collaborative decisions against the initial problem, returning feedback to expert recruitment when incorrect and accepting satisfactory decisions as final answers.The current configuration uses an agent evaluator; human evaluation is proposed for future experiments.Source: A CONFIGURATIONS OF THE EXPERIMENTS · p. 17

B EXPERIMENT DETAILS FOR MULTI-AGENT TOOL USING · B.1 SETUPS

AGENTVERSE’s collaborative tool-use setup combines horizontal, sequential task discussions and subtask summarization with iterative ReAct execution. Agents use OpenAI function calls and stop execution upon submitting a conclusion or reaching ten iterations.

B.1 SETUPS

  • Agents recruited during Expert Recruitment discuss the assigned task using a horizontal communication structure.Source: §B.1 · p. 17
  • An additional summarizer agent consolidates the discussed subtasks for each participant after the discussion concludes.Source: §B.1 · p. 17
  • Agents communicate in a predetermined sequential order during collaborative discussions.Source: §B.1 · p. 17
  • ReAct execution alternates agents’ written thoughts with tool calls, returning tool responses to agents for subsequent reasoning and action.Source: §B.1 · p. 17
  • 10 iterations is the execution limit; the loop ends earlier if the agent submits its conclusion.Source: §B.1 · p. 17

B EXPERIMENT DETAILS FOR MULTI-AGENT TOOL USING · B.1 SETUPS

  • OpenAI’s function calling handles tool invocation.Unless otherwise specified, the implementation follows the standard procedures used in the other experiments.Source: §B.1 · p. 17

B.2 RESULTS

Across the illustrated tasks, AgentVerse often covered requested components omitted by a single ReAct agent, although some tasks showed comparable coverage or incomplete answers from both.

Mathematical games

  • AgentVerse explained and coded the 24-Point Game and supplied similar games, whereas the single ReAct agent omitted similar games.Both responses reported testing their code; AgentVerse additionally described Make a Number and Math Safe.Source: §B.2 · pp. 17–18

Party planning and winter fashion

  • Both systems covered party venues, weather, activities, and winter-fashion shopping; AgentVerse offered more party detail and separated men's and women's fashion trends.Its party response expanded venue and activity suggestions and mentioned icebreakers for guests unfamiliar with one another.Source: §B.2 · pp. 18, 21

Trending topics

  • AgentVerse covered trending music, movies, books, and products, whereas the single ReAct agent returned products and could not find trending books.AgentVerse cautioned that popularity could vary by region and consumer demographic.Source: §B.2 · pp. 18–19

Japan travel planning

  • AgentVerse supplemented Japanese tourist attractions with intercity transportation and historical weather guidance, while the single ReAct agent supplied only attractions.It recommended Shinkansen travel and offered alternatives, but substituted historical averages for an exact next-month forecast and advised checking forecasts closer to departure.Source: §B.2 · p. 19

Date planning and DIY projects

  • For date planning and DIY projects, AgentVerse added missing theaters and materials, respectively, while ReAct supplied restaurants, movies, and project suggestions.AgentVerse named a cinema alongside its restaurant and movie recommendations, and listed project-specific tools for painting, cabinet pulls, television mounting, and fire-pit construction.Source: §B.2 · pp. 20–21

Book-club research

  • Neither system identified the requested bestselling books or summaries; AgentVerse recovered buying and borrowing platforms, while the single ReAct agent failed the task.AgentVerse listed services including OverDrive, the Amazon Kindle Store, and other purchasing or lending platforms.Source: §B.2 · p. 20

Car comparison

  • AgentVerse covered charging, safety, and after-sales service across Ford, Toyota, and Tesla, whereas ReAct covered only Tesla; AgentVerse nevertheless lacked Tesla pricing.AgentVerse supplied starting prices for Ford and Toyota, while the single-agent response supplied Tesla pricing.Source: §B.2 · pp. 20–21

Fitness planning

  • 24.22 was the BMI reported by both systems; AgentVerse supplied explicit daily calorie intake and staged workouts, whereas ReAct specified a calorie-deficit target.AgentVerse proposed 1400 to 1600 calories per day and a 4-week program followed by a 12-week program, with workouts lasting approximately 20-22 minutes.Source: §B.2 · pp. 21–22

C DETAILS OF THE EXPERIMENTS ON MINECRAFT · D PROMPTS

The Minecraft experiments use manually assigned players, consensus-based coordination, Voyager task execution, and inventory/subtask feedback. The prompts section locates the Section 3 prompts in Figures 7–11.

C DETAILS OF THE EXPERIMENTS ON MINECRAFT

  • Automated expert recruitment is bypassed: each agent is manually assigned as an experienced Minecraft player to support consistent teams across multiple rounds.The setup reflects gameplay requiring intricate communication and coordination.Source: C DETAILS OF THE EXPERIMENTS ON… · p. 22
  • Agents communicate horizontally in a predetermined sequence, independently formulating plans and appending [END] when they perceive consensus on task assignments.This communication paradigm supports multiplayer gameplay rather than requiring a single shared solution.Source: C DETAILS OF THE EXPERIMENTS ON… · p. 22
  • After consensus, an auxiliary agent extracts each agent’s assignment from the full communication record and passes it to the corresponding Voyager agent.Source: C DETAILS OF THE EXPERIMENTS ON… · p. 22
  • Several Voyager agents share a Minecraft environment, decomposing assigned tasks into subtasks that they attempt sequentially.The implementation uses the official-repository checkpoint and GPT-4-0314 backbone, consistent with Wang et al. (2023a).Source: C DETAILS OF THE EXPERIMENTS ON… · p. 22
  • Execution ends when an agent completes its task or all agents reach five attempts; evaluation uses inventories and completed or failed subtasks as feedback.Source: C DETAILS OF THE EXPERIMENTS ON… · p. 22

D PROMPTS

  • The prompts used in Section 3 appear in Figures 7–11, with Commongen-Challenge assigned to Figure 10.Source: D PROMPTS · p. 22

E LIMITATION AND FUTURE WORK

AGENTVERSE’s future directions include more capable agents for harder tasks, better multi-party communication, and leveraging positive emergent behaviors while mitigating harmful ones.

More Capable Agents and More Challenging Scenarios

  • More capable, multimodal agents could extend AGENTVERSE beyond its current GPT-4-based agents to more complex real-world tasks.ChatGPT’s voice and image capabilities illustrate perceptual advances that could enable agents to see, hear, and speak.Source: E LIMITATION AND FUTURE WORK · p. 23

Multi-party Communication Among Agents

  • Multi-party collaboration requires agents to decide autonomously when to speak and whom to address, creating communication difficulties during AGENTVERSE’s collaborative decision-making.Proposed directions include better communication-management mechanisms and perceptual-aware LLMs that autonomously interact with their environments, including other agents.Source: E LIMITATION AND FUTURE WORK · p. 23

Leverage Emergent Behaviors and Mitigate Safety Issues

  • Future work should explore leveraging positive emergent behaviors to improve efficiency and effectiveness while mitigating harmful behaviors.The paper identifies both positive and harmful emergent behaviors during agent collaboration.Source: E LIMITATION AND FUTURE WORK · p. 23

F EXAMPLES OF THE CASE STUDIES

The case studies illustrate single-agent and multi-agent software development and consulting, alongside collaborative tool use and Minecraft crafting.

Tool Utilization and Minecraft

Software Development

  • 9 for Completeness, 9 for Functionality, 8 for Readability, and 8 for Robustness appear in one calculator evaluation.Another evaluation reports Completeness:8, Functionality:7, Readability:7, and Robustness:7, requesting safer expression evaluation, more comments, and more exception handling. The examples juxtapose single-agent and multi-agent calculator development and generated code; the higher-scoring evaluation praises structure and explanatory comments.Source: F EXAMPLES OF THE CASE STUDIES · pp. 23, 29, Figure 12 · p. 29, Figure 13 · p. 30

Consulting in Horizontal Structure

Consulting in Vertical Structure

Related papers

Loading 2308.10848v3…