Cost-Effective Communication: An Auction-based Method for Language Agent Interaction

TL;DR

This work tackles the inefficiency of unrestricted inter-agent communication in LLM-based MAS by introducing DALA, a centralized auction framework where speaking rights are allocated to agents via value-density bids learned through MAPPO. By treating communication bandwidth as a scarce resource and implementing a combinatorial VCG auction with a tiered content policy, DALA achieves state-of-the-art results across seven reasoning benchmarks with substantially reduced token usage. The approach fosters emergent strategic silence, guiding agents to speak only when their messages provide meaningful marginal value within budget constraints. Overall, DALA demonstrates that resource-aware, market-based communication can balance performance with efficiency in multi-agent LLM systems.

Abstract

Multi-agent systems (MAS) built on large language models (LLMs) often suffer from inefficient "free-for-all" communication, leading to exponential token costs and low signal-to-noise ratios that hinder their practical deployment. We challenge the notion that more communication is always beneficial, hypothesizing instead that the core issue is the absence of resource rationality. We argue that "free" communication, by ignoring the principle of scarcity, inherently breeds inefficiency and unnecessary expenses. To address this, we introduce the Dynamic Auction-based Language Agent (DALA), a novel framework that treats communication bandwidth as a scarce and tradable resource. Specifically, our DALA regards inter-agent communication as a centralized auction, where agents learn to bid for the opportunity to speak based on the predicted value density of their messages. Thus, our DALA intrinsically encourages agents to produce concise, informative messages while filtering out low-value communication. Extensive and comprehensive experiments demonstrate that our economically-driven DALA achieves new state-of-the-art performance across seven challenging reasoning benchmarks, including 84.32% on MMLU and a 91.21% pass@1 rate on HumanEval. Note that this is accomplished with remarkable efficiency, i.e., our DALA uses only 6.25 million tokens, a fraction of the resources consumed by current state-of-the-art methods on GSM8K. Further analysis reveals that our DALA cultivates the emergent skill of strategic silence, effectively adapting its communication strategies from verbosity to silence in a dynamical manner via resource constraints.

Figures (3)

• Figure 1: An overview of our DALA. Specifically, an Actor Network generates candidate messages, and a Critic Network computes their value density ($\rho_{i}$) based on utility ($V_{i}$) and cost. This density serves as a bid in a budget-constrained VCG auction that allocates speaking rights. The entire actor-critic system is trained with MAPPO to maximize a reward signal combining task success and auction costs, thereby learning an efficient communication strategy.
• Figure 2: The learning curve of the average predicted value ($v_i$) on MMLU, showing the agent's internal value network rapidly learns to assign high value to critical information while devaluing non-critical information, with the valuation gap stabilizing as training converges around epoch 160.
• Figure 3: A comparison of agent communication strategy distributions under two total budget constraints across MMLU domains: $1 \times 10^5$ tokens (bottom) and $1 \times 10^6$ tokens (top). This visual evidence highlights the agents' dynamic adaptation to resource availability.