Multi-agent systems powered by large language models: applications in swarm intelligence

Probably not a substantial claim: may not pass S3 (load-bearing).

The language model offered feedback by highlighting misunderstandings or ambiguities in the prompts.

“Throughout this iterative process, the language model offered valuable feedback by highlighting misunderstandings or ambiguities in the prompts.”

May combine two claims.

In both example applications, agents receive localized information about their environment and act based on LLM-generated decisions.

“In both cases, agents receive localized information about their environment and act based on LLM-generated decisions.”

Classified as probably not a substantial claim.

More explicit rules were essential for consistent and realistic ant behavior.

“Following the initial evaluation, it became clear that more explicit rules were essential for consistent and realistic ant behavior.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The feedback from the language model was instrumental in refining the prompts to enhance the ants' performance.

“This feedback was instrumental in refining the prompts to enhance the ants' performance.”

May combine two claims.

Explicitly stating the ants' status in the prompt improved decision-making by providing clear context, leading to better execution of actions like picking up and dropping food.

“Explicitly stating the ants' status improved decision-making by providing clear context, leading to better execution of actions like picking up and dropping food.”

May combine two claims.

Instructing ants to release pheromones while carrying food led to stronger trails and improved the efficiency of other ants in locating food sources.

“This led to stronger trails and improved the efficiency of other ants in locating food sources.”

May combine two claims.

Ants became more consistent in moving toward stronger cues, enhancing their navigation and foraging efficiency.

“Ants became more consistent in moving toward stronger cues, enhancing their navigation and foraging efficiency.”

May combine two claims.

This eliminated unnecessary returns and improved foraging efficiency.

“This eliminated unnecessary returns and improved foraging efficiency.”

May combine two claims.

From Prompt 5 onward, ants became more adept at following pheromone and nest scent gradients, leading to a more organized foraging behavior.

“From Prompt 5 onward, ants became more adept at following pheromone and nest scent gradients, leading to a more organized foraging behavior.”

May combine two claims.

Ants continued to exhibit limited exploration when not carrying food, tending to stay near the nest rather than venturing into new areas or effectively following pheromone trails.

“Ants continued to exhibit limited exploration when not carrying food, tending to stay near the nest rather than venturing into new areas or effectively following pheromone trails.”

The quote only partly states this claim.

Expanding the rotation options to include 'random' led ants to demonstrate more varied and unpredictable movement patterns during exploration.

“ants demonstrated more varied and unpredictable movement patterns during exploration”

May combine two claims.

A noticeable bias concerning the rotation remained, particularly around the nest, indicating that the randomness was not functioning as efficiently as intended.

“a noticeable bias concerning the rotation remained, particularly around the nest, indicating that the randomness was not functioning as efficiently as intended”

May combine two claims.

Introducing a directive for proactive exploration improved exploration, with ants venturing further from the nest and discovering food sources in fewer simulation steps.

“This approach improved exploration, with ants venturing further from the nest and discovering food sources in fewer simulation steps.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

They effectively moved away from the nest and searched a wider area, increasing their chances of encountering food sources more quickly and efficiently.

“They effectively moved away from the nest and searched a wider area, increasing their chances of encountering food sources more quickly and efficiently.”

May combine two claims.

Adding a directive restricting the response format to a JSON object successfully constrained the output, making it more reliable and compatible with the simulation.

“This modification successfully constrained the output, making it more reliable and compatible with the simulation.”

Classified as probably not a substantial claim. May combine two claims.

Some ants released pheromones unexpectedly, while others exhibited random movement patterns.

“Some ants released pheromones unexpectedly, while others exhibited random movement patterns.”

May combine two claims.

When generating a reasoning process before outputting the final heading, the LLM produced correct answers; errors occurred when providing only the numerical result.

“When generating a reasoning process before outputting the final heading, the LLM produced correct answers. However, errors occurred when providing only the numerical result.”

May combine two claims.

Although central tendencies may be similar, the variance and distributional shape of heading differences differ between Hybrid and NetLogo.

“In summary, these results suggest that although central tendencies may be similar, the variance and distributional shape differ”

May combine two claims.

The Kruskal-Wallis test yields a p-value close to zero, signifying statistically significant differences in the overall distributions of the groups.

“Furthermore, the Kruskal-Wallis test yields a p-value close to zero as well, signifying statistically significant differences in the overall distributions of the groups.”

May combine two claims.

Hybrid starts returning food to the nest at around 20 simulation steps, whereas LLM and NetLogo begin this process at about 40 steps.

“Hybrid starts returning food to the nest at around 20 simulation steps, whereas LLM and NetLogo begin this process at about 40 steps.”

May combine two claims.

Rule-based birds exhibit the highest number of neighbors, while LLM-guided birds display the lowest number.

“As expected, rule-based birds exhibit the highest number of neighbors, while the LLM-guided birds display the lowest number”

May combine two claims.

The standard deviation for NetLogo is around 20, whereas LLM displays a much lower standard deviation of about 7.

“the standard deviation for NetLogo is around 20, whereas LLM displays a much lower standard deviation of about 7”

May combine two claims.

The Levene and Brown-Forsythe tests for the LLM vs. NetLogo group yielded p-values around 0.0045, indicating a significant difference in variance.

“The Levene and Brown-Forsythe tests for the LLM vs. NetLogo group yielded p-values around 0.0045, indicating a significant difference in variance”

May combine two claims.

Other tested LLMs often exhibited unreliable and erratic behavior, including ants randomly dropping food and failing to demonstrate expected self-organizing capabilities.

“Other tested models often exhibited unreliable and erratic behavior, including ants randomly dropping food and failing to demonstrate expected self-organizing capabilities.”

May combine two claims.

While heading convergence sometimes lagged behind purely rule-based simulations, the resulting formations remained visually coherent.

“While heading convergence sometimes lagged behind purely rule-based simulations, the resulting formations remained visually coherent.”

May combine two claims.

The Hybrid model outperforms the other two variants, collecting an average of approximately 95 units of food with a standard deviation of about 12.

“the Hybrid model outperforms the other two variants, collecting an average of approximately 95 units of food with a standard deviation of about 12”

May combine two claims.

Levene's test and Brown-Forsythe test return extremely low p-values, indicating that the assumption of homogeneity of variances is violated.

“Specifically, both Levene's test and the Brown-Forsythe test return extremely low p-values, indicating that the assumption of homogeneity of variances is violated.”

May combine two claims.

The rule-based birds of the original NetLogo model show somewhat lower heading differences than the two bird types of Hybrid, which show similar evolution of heading differences.

“While the two bird types of Hybrid show a similar evolution of the heading differences, the rule-based birds of the original NetLogo model show somewhat lower heading differences.”

May combine two claims.

Explicitly instructing the model to select the shortest path to the target heading improved flocking behavior significantly, with LLM-driven agents forming larger, more stable flocking clusters, achieving performance comparable to the original rule-based NetLogo model.

“By explicitly instructing the model to select the shortest path to the target heading, flocking behavior improved significantly. The LLM-driven agents formed larger, more stable flocking clusters, achieving performance comparable to the original, rule-based NetLogo model.”

May combine two claims.

The Cohen's d effect size for LLM vs. Hybrid is approximately 0.2, suggesting a small yet meaningful effect.

“Additionally, Cohen's d effect size is approximately 0.2, suggesting a small yet meaningful effect.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A notable outlier is observed in the context of NetLogo and food patch 2, where an ant required 720 steps to find food.

“A notable outlier is observed in the context of NetLogo and food patch 2, where an ant required 720 steps to find food.”

Probably not a substantial claim: may not pass S4 (based).

Even small changes in the prompts, such as specifying the angle to rotate or how to handle conflicting signals, can significantly influence emergent group-level patterns.

“Even small changes in the prompts, such as specifying the angle to rotate or how to handle conflicting signals, can significantly influence emergent group-level patterns.”

Classified as probably not a substantial claim. May combine two claims.

The integration of LLMs into agent-based simulations introduces significant computational overhead compared to traditional rule-based implementations, with each agent-LLM interaction requiring network communication and remote inference that substantially increases simulation time.

“The integration of LLMs into agent-based simulations introduces significant computational overhead compared to traditional rule-based implementations, with each agent-LLM interaction requiring network communication and remote inference that substantially increases simulation time.”

May combine two claims.

The Cohen's d value of 0.006 reflects a small effect size, suggesting nearly identical group means for LLM vs. NetLogo.

“the Cohen's d value of 0.006 reflects a small effect size, suggesting nearly identical group means”

May combine two claims.

LLM-driven birds generally stayed slightly farther from the flock center and experienced fewer collisions than their rule-based counterparts.

“However, LLM-driven birds generally stayed slightly farther from the flock center and experienced fewer collisions than their rule-based counterparts”

May combine two claims.

In our specific experiments, each step involving an LLM-driven agent interaction typically required processing times on the order of seconds, whereas conventional rule-based simulations executed agent interactions within milliseconds or less per simulation step.

“In our specific experiments, we observed that each step involving an LLM-driven agent interaction typically required processing times on the order of seconds due to network latency and model response times, whereas conventional rule-based simulations executed agent interactions within milliseconds or less per simulation step.”

Probably not a substantial claim: may not pass S4 (based).

In the ant foraging scenario, complete depletion of food resources typically occurs around 2000 steps.

“particularly since complete depletion of food resources typically occurs around 2000 steps.”

May combine two claims.

Comparing the resulting headings with those produced by a rule-based model, some calculated headings were numerically similar, but they often pointed in opposite directions.

“While some calculated headings were numerically similar, they often pointed in opposite directions.”

May combine two claims.

While some flocking behavior emerged after the JSON restriction, it was inconsistent; small clusters formed briefly, but overall alignment and coherence were weaker than expected.

“while some flocking behavior emerged, it was inconsistent. Small clusters formed briefly, but overall alignment and coherence were weaker than expected.”

May combine two claims.

In the bird flock simulation, collective behaviors stabilized after approximately 500 steps, with cluster formations remaining relatively unchanged in subsequent simulation steps.

“In the bird flock simulation, we empirically observed that collective behaviors stabilized after approximately 500 steps, with cluster formations remaining relatively unchanged in subsequent simulation steps.”

May combine two claims.

Explicitly specifying the compass convention for heading calculations improved flocking behavior, with larger clusters formed compared to previous iterations.

“By explicitly specifying the compass convention for heading calculations, flocking behavior improved. Larger clusters formed compared to previous iterations.”

May combine two claims.

Introducing the 'rationale' key significantly improved flocking behavior, with agents demonstrating more consistent heading adjustments, enhancing the emergence of flocking dynamics.

“Introducing the “rationale” key significantly improved flocking behavior. The agents demonstrated more consistent heading adjustments, enhancing the emergence of flocking dynamics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The final prompt used in the experiments resulted from several iterations in a trial-and-error process to ensure the LLM could effectively understand the environment and rules and execute the expected tasks accordingly.

“The final prompt used in our experiments resulted from several iterations in a trial-and-error process to ensure the LLM could effectively understand the environment and rules and execute the expected tasks accordingly.”

• AssumptionRests on an assumption: The LLM can effectively understand the environment and rules and execute the expected tasks after prompt iterations.

Classified as probably not a substantial claim. May combine two claims.

The agents are designed to follow explicit, rule-based instructions derived from the LLM, ensuring that their actions are predictable and consistent with predefined conditions.

“The agents are designed to follow explicit, rule-based instructions derived from the LLM, ensuring that their actions are predictable and consistent with predefined conditions.”

• AssumptionRests on an assumption: The LLM's ability to process complex, context-rich data allows it to infer and propose actions that adhere to predefined rules or leverage general behavioral principles.

Classified as probably not a substantial claim.

Knowledge-driven prompts enable more adaptive and flexible agent behaviors by relying on the LLM's inherent understanding of broader behavioral concepts and principles.

“Knowledge-driven prompts, on the other hand, rely on the LLM's inherent understanding of broader behavioral concepts and principles.”

• Also stated in: § Materials and methods · plain

  “These prompts are less rigid and provide the LLM with general guidelines, enabling more adaptive and flexible agent behaviors.”

• AssumptionRests on an assumption: The LLM has inherent understanding of complex concepts such as flocking dynamics and self-organization.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

This work is one of the first to use LLMs as decentralized behavioral engines for non-human and non-verbal, swarm-like agents.

“This is one of the first works to use LLMs as decentralized behavioral engines for non-human and non-verbal, swarm-like agents.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

There is a small chance that the LLM may occasionally generate responses that deviate from the specified rules.

“Nevertheless, there is still a small chance that the LLM may occasionally generate responses that deviate from the specified rules.”

Classified as probably not a substantial claim.

The iterative process of the toolchain ensures that agent behaviors continuously respond to evolving environmental conditions and LLM feedback, fostering emergent behaviors and adaptive responses.

“The iterative process ensures that agent behaviors continuously respond to evolving environmental conditions and LLM feedback, fostering emergent behaviors and adaptive responses.”

• AssumptionRests on an assumption: The LLM's ability to process complex, context-rich data allows it to infer and propose actions that adhere to predefined rules or leverage general behavioral principles.

May combine two claims.

This change increased the likelihood of ants finding food, as they ventured further from the nest rather than lingering nearby.

“This change increased the likelihood of ants finding food, as they ventured further from the nest rather than lingering nearby.”

Classified as probably not a substantial claim.

Structured prompts ensure that the ants' responses are clearly defined and predictable, enabling systematic analysis of their behavior.

“These structured prompts ensure that the ants' responses are clearly defined and predictable, enabling systematic analysis of their behavior.”

• AssumptionRests on an assumption: The LLM's ability to process complex, context-rich data allows it to infer and propose actions that adhere to predefined rules or leverage general behavioral principles.

Probably not a substantial claim: may not pass S3 (load-bearing).

Setting the temperature to 0.0 ensures deterministic results based on the provided inputs.

“The default API parameters are used, with the exception of the temperature, which is set to 0.0 to ensure deterministic results based on the provided inputs.”

May combine two claims.

There is a statistically significant and practically noticeable difference between the LLM and Hybrid groups.

“This analysis reveals a statistically significant and practically noticeable difference between the LLM and Hybrid groups.”

Probably not a substantial claim: may not pass S3 (load-bearing).

When using an external API, much of the computational burden is offloaded to the AI service provider, alleviating the direct computational cost from the simulation user's perspective.

“It is important to clarify, however, that when using an external API (such as the OpenAI API), much of this computational burden is offloaded to the AI service provider, thereby alleviating the direct computational cost from the simulation user's perspective.”

May combine two claims.

Although significantly more computationally intensive per agent-step, the method remains practically feasible for smaller-scale exploratory simulations and conceptual validations.

“Thus, although significantly more computationally intensive per agent-step, the method remains practically feasible for smaller-scale exploratory simulations and conceptual validations, as demonstrated in this paper.”

May combine two claims.

Although the experiments utilized a single external model, the methodological approach is fully extendable and compatible with a variety of language models.

“Although the experiments presented in this study utilized a single external model, our methodological approach is fully extendable and compatible with a variety of language models.”

May combine two claims.

Stateless LLM decision-making guarantees that the model consistently acts on relevant information but also necessitates highly detailed prompts.

“This guarantees that the model consistently acts on relevant information but also necessitates highly detailed prompts.”

Probably not a substantial claim: may not pass S2 (checkable).

This work underscores the potential for advanced language models, guided by carefully designed prompts, to enrich or even extend the capabilities of traditional agent-based models.

“Ultimately, this work underscores the potential for advanced language models, guided by carefully designed prompts, to enrich or even extend the capabilities of traditional agent-based models, offering new perspectives on swarm intelligence, self-organization, and emergent behaviors.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The methodological framework is not inherently tied to GPT-4o or the OpenAI API; it was explicitly designed to ensure adaptability and generality, allowing straightforward integration with alternative LLMs.

“However, it is crucial to emphasize that our framework is not inherently tied to GPT-4o or the OpenAI API. Rather, we explicitly designed our methodological framework to ensure adaptability and generality, allowing straightforward integration with alternative LLMs, including open-source and locally deployed solutions.”

May combine two claims.

LLMs can support swarm-like behaviors: guiding ants to locate and retrieve food by following pheromone trails, and prompting "birds" to coordinate alignment according to core flocking principles.

“In both cases, our experiments demonstrated how LLMs can support swarm-like behaviors: guiding ants to locate and retrieve food by following pheromone trails, and prompting “birds” to coordinate alignment according to core flocking principles.”

Probably not a substantial claim: may not pass S2 (checkable).

The toolchain integrating NetLogo simulations with LLM interactions via Python extension and the OpenAI API offers a practical contribution in terms of accessibility and pedagogical potential.

“Additionally, in line with our research objectives, the toolchain we present, integrating NetLogo simulations with LLM interactions via Python extension and the OpenAI API, offers a practical contribution in terms of accessibility and pedagogical potential.”

May combine two claims.

Concerning food patch 1, the models exhibit a higher number of outliers, which can be attributed to the ants' initial exploration of the environment before encountering the food.

“Notably, concerning food patch 1, the models exhibit a higher number of outliers, which can be attributed to the ants' initial exploration of the environment before encountering the food.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Our findings highlight the central role of iterative prompt tuning in shaping how LLMs behave within multi-agent environments.

“Our findings highlight the central role of iterative prompt tuning in shaping how LLMs behave within multi-agent environments.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Accessing commercial LLM APIs like GPT-4o incurs token-based costs that scale with simulation complexity, agent population, and runtime duration.

“Additionally, accessing commercial LLM APIs like GPT-4o incurs token-based costs that scale with simulation complexity, agent population, and runtime duration.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Utilizing GPT-4o through the OpenAI API inherently introduces token-based costs and dependencies associated with external model access, which may affect simulation scalability, cost-efficiency, and reliability.

“Second, utilizing GPT-4o through the OpenAI API inherently introduces token-based costs and dependencies associated with external model access, which may affect simulation scalability, cost-efficiency, and reliability.”

May combine two claims.

There are no practical or statistical differences observed between the LLM and NetLogo groups regarding the outcome.

“In summary, there are no practical or statistical differences observed between the LLM and NetLogo groups regarding the outcome.”

May combine two claims.

These results confirm that LLMs can serve as flexible engines for agent behaviors that align with swarm principles, offering adaptive, context-driven responses.

“Together, these results confirm that LLMs can serve as flexible engines for agent behaviors that align with swarm principles, offering adaptive, context-driven responses.”

May combine two claims.

Overall, the LLM-driven agents performed comparably to their fully rule-based counterparts, but they sometimes displayed notable differences in how they interpreted and prioritized local cues when relying on text-based decision-making.

“Overall, the LLM-driven agents performed comparably to their fully rule-based counterparts, but they sometimes displayed notable differences in how they interpreted and prioritized local cues when relying on text-based decision-making.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

Incorporating LLMs into NetLogo facilitates the study of cognitively richer and more adaptive agent behaviors, while also lowering the barrier to entry for researchers, educators, and students.

“By incorporating LLMs into this familiar and user-friendly environment, our approach facilitates the study of cognitively richer and more adaptive agent behaviors, while also lowering the barrier to entry for researchers, educators, and students interested in experimenting with advanced AI techniques in agent-based modeling.”

Probably not a substantial claim: may not pass S2 (checkable).

Careful attention to the length, structure, and clarity of the prompts is essential to achieve consistent, context-aware responses.

“Careful attention to the length, structure, and clarity of the prompts is essential to achieve consistent, context-aware responses.”