Research map / Multi-agent systems and games / LLM agents

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

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DALA, a multi-agent LLM framework where agents bid for speaking rights in a budget-constrained VCG combinatorial auction based on message value density, trained with MAPPO to balance task reward and communication cost.

Problem

Multi-agent LLM systems use inefficient "free-for-all" communication that incurs exponential token costs and low signal-to-noise ratio.

Approach

Treats communication bandwidth as a scarce tradable resource: agents bid for the right to speak using the normalized value density of their candidate messages, with winners chosen by a budget-constrained combinatorial auction and charged VCG payments that enter the MAPPO reward.

Key results

• DALA reports 84.32% on MMLU, 96.18% on GSM8K, and 91.21% pass@1 on HumanEval, the best among the compared methods in the main table.
• On GSM8K, DALA uses about 6.25 million tokens versus about 14 million for DyLAN and 26 million for PHP in the efficiency table.
• Ablation removing value learning drops MMLU accuracy from 84.32 to 75.77 and GSM8K from 96.18 to 86.35.

Surprising findings

• Under a low token budget, agents in STEM rely mostly on keywords and silence, while under a high budget they favor full-text messages, which the authors interpret as emergent strategic silence.

Limitations

• WDP is NP-hard in general and relies on dynamic programming, which may not scale to large agent groups.
• Reported numbers are inconsistent across tables (e.g., GSM8K 96.18 vs 96.2, MMLU tokens 1.81e5 vs 3.80e5), and the paper's text cites an appendix not included.

Experimental setup

Datasets: MMLU GSM8K MultiArith SVAMP AQuA MATH-500 HumanEval

Benchmarks: MMLU MultiArith GSM8K SVAMP AQuA HumanEval MATH-500

Metrics: accuracy pass@1 token consumption

Baselines: Vanilla CoT Complex CoT Self-Consistency PHP LLM-Debate DyLAN AgentPrune-R

Builds on: MAPPO AgentPrune LLM-Debate DyLAN PHP

Closest papers by meaning

• Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems (ICLR 2025 · similarity 0.89)
• Scaling Small Agents Through Strategy Auctions (ICML 2026 · similarity 0.89)
• More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.88)
• Communicating Activations Between Language Model Agents (ICML 2025 · similarity 0.87)
• A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive Reliability (NeurIPS 2026 · similarity 0.87)
• Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.87)