Source-linked AI summary
Exploring Collaboration Mechanisms for LLM Agents: A Social Psychology View
Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu, Bryan Hooi, Shumin Deng
TL;DR
The paper asks whether multi-agent LLMs can develop collaborative intelligence and social behaviors in machine societies. It builds societies with different traits and thinking patterns, then evaluates their strategies across three datasets. The results identify strategy-dependent performance and efficiency patterns alongside human-like conformity and consensus behaviors.
Problem
The paper addresses the limited understanding of how LLMs manifest social intelligence and collaboration within machine societies.
Method
The authors evaluate four three-agent societies with easy-going or overconfident traits and debate or reflection strategies across three datasets.
Results
Collaborative-strategy performance varies, uniform within-round thinking improves efficiency, and LLM agents exhibit conformity and consensus-reaching behaviors.
Takeaways & Limitations
The findings support designing multi-agent systems through effective collaboration strategies rather than relying only on larger numbers of LLMs or agents.
Takeaways & Limitations
The study does not explore societies whose agents use different LLMs and relies on exhaustive search rather than agents autonomously selecting strategies.
Abstract
from arXiv · showhide
As Natural Language Processing (NLP) systems are increasingly employed in intricate social environments, a pressing query emerges: Can these NLP systems mirror human-esque collaborative intelligence, in a multi-agent society consisting of multiple large language models (LLMs)? This paper probes the collaboration mechanisms among contemporary NLP systems by melding practical experiments with theoretical insights. We fabricate four unique `societies' comprised of LLM agents, where each agent is characterized by a specific `trait' (easy-going or overconfident) and engages in collaboration with a distinct `thinking pattern' (debate or reflection). Through evaluating these multi-agent societies on three benchmark datasets, we discern that certain collaborative strategies not only outshine previous top-tier approaches, but also optimize efficiency (using fewer API tokens). Moreover, our results further illustrate that LLM agents manifest human-like social behaviors, such as conformity and consensus reaching, mirroring foundational social psychology theories. In conclusion, we integrate insights from social psychology to contextualize the collaboration of LLM agents, inspiring further investigations into the collaboration mechanism for LLMs. We commit to sharing our code and datasets\footnote{\url{https://github.com/zjunlp/MachineSoM}.}, hoping to catalyze further research in this promising avenue.
1 Introduction
The paper studies whether multi-agent LLM societies can exhibit collaborative intelligence and social behaviors, using social psychology to frame their interactions. Across three datasets, it examines how traits and thinking-pattern strategies affect performance and efficiency.
- The test-bed combines four societies of three agents, easy-going or overconfident traits, debate or reflection patterns, and multiple collaborative strategies across three datasets.
- Collaborative strategies differ significantly in performance, with substantive debate enhancing collaboration while societies with different traits show no clear performance differences.
- Uniform thinking patterns within each collaboration round improve efficiency, while increasing agent numbers or rounds does not consistently improve outcomes.
- LLM agents display human-like social behaviors, including conformity, majority-rule thinking, and consensus reaching, paralleling social psychology theories.
- The findings motivate collaboration designs that move beyond simply scaling LLMs and agents toward more socially aware multi-agent systems.
- The paper investigates how and to what extent LLMs manifest social intelligence through collaboration in multi-agent societies.
2 Explore Collaboration Mechanisms with Multiple LLM Agents
The paper formalizes machine societies by combining agent traits, thinking patterns, and multi-round collaborative strategies. It evaluates these configurations on multiple reasoning and decision-making tasks using accuracy, variability, and token-efficiency measures.
- Society Simulation: Collaborative strategies are permutations of thinking patterns across multiple rounds, including uniform patterns and configurations where one agent differs from others.
- Society Simulation: Each machine society contains n LLM agents with overconfident or easy-going traits and debate or reflection as available thinking patterns.
- Society Simulation: The primary study constructs four three-agent societies ranging from totally overconfident to totally easy-going configurations.
- Society Simulation: With three agents and three rounds, the simulation evaluates eight possible collaborative strategies and later considers more agents, rounds, and strategy combinations.
- Experimental Settings: Evaluation covers MMLU high-school multiple choice, advanced MATH problems, and Chess Move Validity, using randomly selected subsets for the first two tasks.
- Experimental Settings: Results report mean accuracy and standard deviations over five trials, supplemented by WIN-TIE counts against the continuous-debate baseline.
- Experimental Settings: Table 2 compares strategy and society performance, token costs, and WIN-TIE counts, with color coding identifying best, second-best, and worst strategies.
3 Analysis of Machine Social Collaboration
The study evaluates how society composition, thinking-pattern permutations, agent counts, and collaboration rounds affect multi-agent LLM performance across tasks. Debate-initial or debate-dominant strategies generally perform best, while three agents and three rounds offer an effective efficiency trade-off.
- Society composition: Three-agent societies with varied traits show no pronounced accuracy differences under the same collaborative strategy.Significance tests further indicate that distinct three-agent societies play an indistinctive role in shaping performance.
- Thinking-pattern permutations: Debate-initial and debate-dominant thinking-pattern strategies generally outperform alternatives across datasets and models.On MMLU, p0p0p1 in S4 reaches 65.2 versus 34.4 for p1p0p0 in S4.
- Task dependence: Continuous debate generally benefits tasks, while debate combined with continuous reflection is superior for difficult tasks such as MATH level 5.The strategies p0p0p0 and p0p0p1 generally perform well, whereas p0p1p1 and p1p1p0 perform better on MATH level 5.
- Agent counts: Three agents are relatively optimal because odd agent counts generally perform better, accuracy differences among odd counts are indistinctive, and more agents reach consensus more often.Different agent counts significantly impact performance, while three agents balance performance and efficiency.
- Collaboration rounds: Three-round collaboration is relatively effective and efficient because performance from three to ten rounds fluctuates without extremely remarkable variation while additional rounds increase cost.The study reports a significance test on the impact of collaboration rounds.
- Thinking-pattern uniformity: Using inconsistent thinking patterns among agents in one round is evaluated against having all agents use the same pattern, with results reported on MATH.The study randomly allocates varied patterns while avoiding rounds in which all agents use the same pattern.
4 Phenomena of Conformity and Consensus Reaching
The analysis finds that LLM-agent collaboration exhibits conformity and consensus-reaching patterns resembling phenomena studied in social psychology. Conformity becomes more prevalent across rounds, but its performance effects vary by model and can weaken or worsen with additional rounds.
- Social-psychology alignment: Machine-society collaboration echoes human societal phenomena including conformity and consensus reaching.The analysis connects these behaviors with social psychology theories and human group dynamics.
- Consensus reaching: Agents can converge on an incorrect answer after being swayed by misguided peers, or converge on a correct answer after society-wide debate.These examples are interpreted as detrimental and beneficial forms of groupthink, respectively.
- Conformity: Conformity is widespread, and its proportion generally increases as collaboration rounds increase.The study analyzes answer-correctness changes at each round of collaboration.
- Conformity: Conformity improves performance on ChatGPT and Qwen 72B but harms performance on LlaMA2 Chat 13B/70B and Mixtral 8×7B.The direction of conformity’s performance effect differs across evaluated language models.
- Conformity: As rounds increase, conformity benefits weaken while conformity detriments strengthen.The ratio difference between True and False answers becomes smaller for benefits and larger for detriments.
- Consensus reaching: Reflection generally increases consensus-cluster quantity, whereas debate is more likely to reach consensus.Fewer consensus clusters indicate easier consensus reaching.
5 Conclusion and Future Work
The conclusion highlights LLM agents’ collaboration capabilities across traits, thinking patterns, and strategies, alongside human-like behaviors connected to social psychology. It calls for deeper study of multi-agent societies and collaboration mechanisms.
- Conclusion: LLM agents show collaboration capabilities across different individual traits, thinking patterns, and collaborative strategies.The conclusion presents these capabilities as evidence of the potential of LLM collaboration mechanisms.
Limitations
The study’s scope is limited by its agent, strategy, society, evaluation, and task configurations.
- Limitations: The study does not explore societies combining multiple different LLMs, adaptive autonomous strategy selection, broader traits, larger societies, or more realistic creative tasks.It uses search over strategies, relatively straightforward configurations, and manual validation or rule-based matching.
Ethics Statement
The research used publicly accessible datasets and no proprietary or confidential information. The authors therefore report no ethical concerns.
- The data came from publicly accessible datasets.
- The study did not use proprietary or confidential information.
- The authors report that the research is free from ethical concerns.
Overview of Appendices
The appendices provide supporting material on takeaways, related work, applications, implementation, social collaboration analyses, alternative backbones, and prompts.
- Appendix A summarizes the paper’s key takeaways.
- Appendices B–D cover related work, potential real-world applications, and implementation details, including setup, evaluation, and agent-collaboration illustrations.
- Appendices E–G analyze machine social collaboration, machine-society settings, conformity, consensus reaching, and group dynamics.
- Appendix H evaluates different backbone LLMs, while Appendix I assesses prompt effectiveness.
A Key Takeaways
The paper recommends selecting collaboration strategies according to task difficulty and using modest, uniform multi-agent settings. Its analyses also identify conformity, consensus, and other human-like social behaviors among LLM agents.
- Strategy Selection: Strategy Selection: Starting or dominating collaboration with debate yields relatively optimal outcomes.
- Strategy Selection: Strategy Selection: Totally-reflection strategy p1p1p1 is generally worst in performance.
- Strategy Selection: Strategy Selection: Difficult tasks favor debate combined with continuous reflection, whereas simple tasks can use self-consistency or reflection.
- Society Settings: Society Settings: Setting agent numbers to 3 is generally advantageous for performance and cost.
- Society Settings: Society Settings: Collaboration rounds are relatively suitable to set as 3 because this is both effective and efficient.
- Social Psychology View: Social Psychology View: Collaboration can produce beneficial or harmful conformity, while increasing rounds decrease conformity’s benefits and increase its detriments.
- Social Psychology View: Social Psychology View: The totally easy-going society is more likely to reach consensus; debate helps consensus reaching, while reflection impedes it.
D.1 Experimental Setup
The experiments compare machine societies and collaborative strategies across benchmark datasets, using controlled settings and majority-vote evaluation against ground truth.
- The study evaluates machine societies on MMLU, Chess Move Validity, and sampled MATH tasks.The experiments sample 50 cases from each dataset.
- Context-length constraints for LlaMA2 Chat 13B and 70B cap collaboration at 4 rounds and 4 agents.
- The number of viable collaborative strategies grows exponentially with each additional round, so the analysis examines 8 representative strategies.Ten rounds would yield 2^10 unique strategies.
- The evaluation selects a unified society answer by majority vote and matches responses against ground truth using comprehensive matching rules.Noncompliant responses and non-unique equivalent answer forms are addressed during evaluation.
- Conformity can improve or harm problem solving by changing a wrong answer to correct or a correct answer to wrong.
E Further Analysis on Machine Social Collaboration (Backbone: ChatGPT)
The ChatGPT analysis tests how societies and collaborative strategies affect performance across three datasets. Collaborative strategy significantly affects accuracy, while society effects are less consistent, and task subject and difficulty change strategy sensitivity.
- The experiments evaluate four society settings and eight collaborative strategies across MATH, MMLU, and Chess Move Validity.
- Collaborative strategy significantly affects accuracy across the ChatGPT experiments.One-way ANOVA tests examine strategy effects across the three datasets.
- Society effects are inconsistent, with significant society differences in 17 of 24 ChatGPT cases for one engine.
- Three-agent societies with different traits show indistinct performance effects across the evaluated tasks.
- Collaborative strategies vary in effectiveness across subjects and difficulty levels.The analysis presents radar maps showing differing task sensitivity to strategies.
F Analysis on Machine Society Settings (Backbone: ChatGPT)
The ChatGPT setting analysis examines agent counts, collaboration rounds, thinking-pattern arrangements, and social behaviors. Performance depends on strategy and agent quantity, while additional rounds increase token use without consistently improving results.
- Different Numbers of Agents: Three agents are relatively optimal, while changing agent numbers significantly correlates with performance.
- Different Rounds of Collaboration: 3-round collaboration is relatively optimal because extra rounds add token consumption without obvious performance gains.
- Consensus Reaching: Consensus clusters decrease after debate but increase after reflection at the same collaboration juncture.
- Other Collaborative Strategies: Uniform thinking patterns within a round improve efficiency, but their accuracy effect varies by dataset.The influence is pronounced on Chess Move Validity and less significant on MMLU and MATH.
- Collaborative Strategies: Debate-initial or debate-dominant strategies mostly outperform self-consistency, although they cost more tokens.
- Collaborative Strategies: Reflection is more likely than debate to produce wavering answers and heightened hallucination risk.
H.1 LlaMA2 Chat 13B
The LlaMA2 Chat 13B analysis extends the society, agent-count, round, strategy, and social-behavior evaluations beyond ChatGPT. It examines how these settings affect accuracy, consensus, conformity, and answer-changing behavior.
- Machine Society Settings: The analysis varies agent numbers, collaboration rounds, and same-versus-hybrid thinking patterns.
- Social Psychology View: The study measures conformity, consensus clusters, and group dynamics through answer-changing behaviors.
- Machine Social Collaboration: The LlaMA2 Chat 13B experiments evaluate eight strategies across three datasets and distinct societies.
- Machine Social Collaboration: The analysis also presents comparative word clouds and trait-based answer retention for overconfident and easy-going societies.
- Statistical Analysis: Figures and significance tests compare collaborative-strategy effects on accuracy for LlaMA2 Chat 13B.