A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration

Recent studies show that collaborating multiple large language model (LLM) powered agents is a promising way for task solving. However, current approaches are constrained by using a fixed number of agents and static communication structures. In this work, we propose automatically selecting a team of agents from candidates to collaborate in a dynamic communication structure toward different tasks and domains. Specifically, we build a framework named Dynamic LLM-Powered Agent Network ($\textbf{DyLAN}$) for LLM-powered agent collaboration, operating a two-stage paradigm: (1) Team Optimization and (2) Task Solving. During the first stage, we utilize an $\textit{agent selection}$ algorithm, based on an unsupervised metric called $\textit{Agent Importance Score}$, enabling the selection of best agents according to their contributions in a preliminary trial, oriented to the given task. Then, in the second stage, the selected agents collaborate dynamically according to the query. Empirically, we demonstrate that DyLAN outperforms strong baselines in code generation, decision-making, general reasoning, and arithmetic reasoning tasks with moderate computational cost. On specific subjects in MMLU, selecting a team of agents in the team optimization stage improves accuracy by up to 25.0% in DyLAN.

Paper

Similar papers

Reviewer sGnL6/10 · confidence 3/52024-05-10

Summary

This paper introduces an approach called DyLAN that incorporates a two-stage paradigm: i) Team Optimization, the temporal feed-forward neural network (T-FFN) that selects most contributary agents for the given task and ii) Task solving, the T-FFN facilitate agents to communicate with each others to solve a bunch of a diverse of tasks. To further enhance the efficiency, DyLAN uses early stopping mechanism to terminate the conversation when most of agents in a single layer agree to have a consistent answer. In the experiment section, the DyLAN is proved to outperform with baselines such as LLM-Blender, LLM Devate, ReAct, and Reflexation, etc, on code generation, decision making, and general reasoning, etc, and shows that DyLAN.

Rating

6

Confidence

3

Ethics flag

1

Reasons to accept

1. Solid Experiments: The paper provides a robust experimental methodology, demonstrating DyLAN's effectiveness across a range of downstream tasks such as coding, general knowledge quetsion answering, math etc. It shows that cost DyLan can enhance the performance of the above mentioned tasks while keep low cost on computation with respect to reasonable times of function call. 2. The idea of DyLAN is novel, it trains a dynamic temporal feedforward network to select the imporatant agents, and the early stopping seems to be effective to reduce the cost on long-term conversation. 3. The ablation study is thorouh.

Reasons to reject

1. **Extended Query on the Selection and Performance of Agents in the Initial Pool**: "Upon close examination of Figure 5 within the manuscript, there appears to be a lack of clarity regarding the roles adopted by the other agents involved in the code generation task. It is noted that the 'programmer' and 'algorithm developer' agents seem to outperform the rest, which might suggest that these two agent roles are disproportionately influential in the task's outcome. This observation raises the question of whether these two agent roles alone suffice for the task at hand. Similarly, in the mathematical reasoning tasks, the 'programmer' and 'mathematician' agents are indicated as the primary contributors, prompting the inquiry of whether including agents with roles such as 'doctor' or 'economist' might be an overcomplication in the context of mathematical tasks. If the initial agent pool is curated for specificity to different tasks, then the inclusion of seemingly irrelevant agents could be considered redundant and may detract from the efficiency of the overall system. Could the authors provide insight into the rationale behind the selection of these varied roles and how their inclusion impacts the performance and generalizability of the DyLAN framework?" 2. Upon reviewing Figures 5 and 6, it is observed that layers with an even number typically assume a critical function, evaluating the responses generated by the preceding odd-numbered layers. This pattern suggests that the effectiveness of the DyLAN framework may be contingent upon the utilization of highly curated prompts tailored to fit each agent's distinct roles and scenarios. The observed phenomenon raises concerns regarding the scalability and flexibility of the system, particularly in scenarios where the diversity of tasks and roles is substantial. It would be beneficial for the authors to address the necessity for such detailed prompt curation and discuss potential strategies to streamline this process, ensuring that the framework remains adaptable across a wide array of tasks without compromising performance. However, I find the paper to be of high quality overall; if the authors can provide clear and satisfactory responses to the queries raised, I would be willing to increase the points.

Reviewer B67Q8/10 · confidence 3/52024-05-11

Summary

In this paper, the author introduce the Dynamic LLM-Agent Network (DyLAN) framework, designed to facilitate collaboration automatically among multiple LLM agents.The author employs T-FNN to dynamically regulate interactions between different agents and Introducing Agent Importance Score to measure each agent's contribution, reducing manual intervention, which is quite innovative in this field. The whole framework includes two stages: "Team Optimization", which identifies and excludes underperforming agents, and "Task solving" performing inference based on the T-FFNs. The author conducted experiments on DyLAN across four tasks: Code Generation, Decision Making, General Reasoning, and Arithmetic Reasoning, comparing it with several previous SOTA methods. DyLAN not only improves performance but also keeps API calls at a relatively low level. To demonstrate the effectiveness of DyLAN, the author conducted lots of experiments and discussions. However, many important findings were included in the appendix. Personally, I believe some crucial discussions such as C.2 and C.8 should be included in the main paper.

Rating

8

Confidence

3

Ethics flag

1

Reasons to accept

1. Using T-FNN to dynamically regulate interactions between different agents and Introducing Agent Importance Score to measure each agent's contribution, reducing manual intervention, which is quite innovative in this field. 2. The idea of introducing Byzantine Consensus theory to achieve early stopping is very interesting and proved to be effictive through ablation study. 3. The author demonstrated the outstanding performance of DyLAN through various experiments and comparisons with current sota methods.

Reasons to reject

I don't have any major concerns that would lead me to reject the submission of this article, although: 1. As I said in summary, some crucial discussions should be included in main paper. 2. The improvement of DyLAN on the AR task is limited compared to other tasks. What might cause this phenomenon?

Questions to authors

1. In the ablation study, I noticed an interesting phenomenon: the early stop has a significant impact on API calls in arithmetic reasoning and general reasoning tasks. Whereas in the decision-making task, both methods have considerable impacts. For code generation, however, neither method seems to have much effect. What reasons might the author attribute to this phenomenon?

Reviewer 2JJf6/10 · confidence 3/52024-05-20

Summary

In this paper, it proposes a new method for dynamically conduct multi-agent collaboration. In incorporates the team operation process and task solving process, which can improve the performance of agent collaborations. The experiments have shown the effectiveness of this method.

Rating

6

Confidence

3

Ethics flag

1

Reasons to accept

Advantage: 1.The problem is practical and the motivation is sensible. It is similar to create a meta agent filter to construct an agent team. Some previous works like MetaGPT also study this problems as well. The techniques sound great, which utilizes methods from graph learni

Reasons to reject

Shortcoming & Question: 1.It seems that there is no error bar in the results of experiments. 2.What about the comparison with MetaGPT? 3.The paper formatting can be improved, such as reducing the empty blocks within it.

Questions to authors

see above

Reviewer sGnL2024-06-05

Response

Thank you for the response. I keep the score (6) unchanged.

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC