TodyComm: Task-Oriented Dynamic Communication for Multi-Round LLM-based Multi-Agent System

TL;DR

TodyComm introduces a task-oriented, dynamic communication framework for multi-round LLM-based MAS that learns round-specific interaction topologies to maximize task utility. It models interactions with a per-round inter-agent DAG $\mathcal{G}_C^t$ and a final decision graph $\mathcal{G}_D$, and uses a gated recurrent network to generate per-agent credits that drive participation and edge construction via policy gradient $u_q(\tau)$-oriented reinforcement learning. The method achieves superior task accuracy and token efficiency across five benchmarks under dynamic adversarial conditions, while maintaining robust adversary detection (ADAcc $>85\%$) and scalability with growing agent counts and budgeted connectivity. Extensive ablations confirm the importance of credit learning, ranking-based graph construction, and rich node features that blend self, neighborhood, and difference information. Overall, TodyComm demonstrates that learning behavior-driven, round-aware topologies yields robust, scalable collaboration in realistic, dynamic MAS settings.

Abstract

Multi-round LLM-based multi-agent systems rely on effective communication structures to support collaboration across rounds. However, most existing methods employ a fixed communication topology during inference, which falls short in many realistic applications where the agents' roles may change \textit{across rounds} due to dynamic adversary, task progression, or time-varying constraints such as communication bandwidth. In this paper, we propose addressing this issue through TodyComm, a \textbf{t}ask-\textbf{o}riented \textbf{dy}namic \textbf{comm}unication algorithm. It produces behavior-driven collaboration topologies that adapt to the dynamics at each round, optimizing the utility for the task through policy gradient. Experiments on five benchmarks demonstrate that under both dynamic adversary and communications budgets, TodyComm delivers superior task effectiveness while retaining token efficiency and scalability.

Figures (8)

• Figure 1: Overview of the training workflow for TodyComm. $\{A_i\}_{i=1}^N$ denote the agents. Given node features $\{f_i^t\}_{i=1}^N$ and hidden states $\{h_i^{t-1}\}_{i=1}^N$, agent-level credits $\{c_i^t\}_{i=1}^{N}$ are generated through GRN, $t \in [2, T+1]$. These credits are used to sample agent participation actions and to deterministically construct the communication graph at each round by prioritizing edges between high-credit agents.
• Figure 2: Training iterations for accuracy (upper row) and adversary detection accuracy (lower row) on two benchmarks.
• Figure 3: Training curves of node feature ablations for TodyComm on MMLU
• Figure 4: G-Safeguard: learning process in dynamic adversarial setting on MMLU under different attack rates. The blue "Training" is evaluated on training data, hence the overfitting.
• Figure 5: Training curves of task accuracy (top row) and adversary detection accuracy (bottom row) for TodyComm across three benchmarks.