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

Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems

ICLR 2025 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 3 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 AgentPrune, which learns low-rank-regularized spatial-temporal communication graph masks in a few dialogue rounds and then one-shot prunes redundant inter-agent messages to cut token cost of LLM multi-agent systems while keeping or improving accuracy and adding robustness to attacks.

Problem

LLM-based multi-agent systems incur large token overhead and economic cost from inter-agent communication, much of which is redundant or even malicious.

Approach

Model the multi-agent system as a spatial-temporal graph (intra-round spatial edges and inter-round temporal edges), define communication redundancy, learn relaxed graph masks with policy gradient and a nuclear-norm low-rank penalty, then apply one-shot magnitude pruning to fix a sparse topology for remaining rounds or queries.

Key results

• AgentPrune with a complete graph (AgentPrune-C) reaches 89.72 average accuracy across six benchmarks using five GPT-4 agents, versus 86.77 for GPTSwarm and 87.02 for PHP.
• Plugged into AutoGen and GPTSwarm, AgentPrune cuts prompt tokens by roughly 27% to 72% (e.g., 71.9% on AutoGen MMLU, 39.4% on GPTSwarm GSM8K) while maintaining or slightly improving accuracy.
• Claims comparable performance on MMLU at $5.6 cost versus $43.7 for state-of-the-art topologies.

Measured effects reported in the paper

Surprising findings

• Randomly pruning 10%-30% of communication connections in MacNet mesh and LLM-Debate structures improved MMLU accuracy by up to 2.83%, indicating substantial redundancy.
• Some multi-agent topologies such as star and tree fail to consistently improve over single-agent baselines, with drops of up to 3.97%.

Limitations

• Experiments rely on closed GPT models via API, so generalization to open-weight models is not shown in the text.
• Multi-agent setups use fixed agent counts (three or five) and specific benchmarks.

Experimental setup

Datasets: MMLU GSM8K MultiArith SVAMP AQuA HumanEval

Benchmarks: MMLU GSM8K MultiArith SVAMP AQuA HumanEval

Metrics: accuracy pass@1 prompt tokens cost (USD)

Baselines: CoT ComplexCoT Self-Consistency LLM-Blender GPTSwarm LLM-Debate PHP DyLAN AutoGen

Builds on: GPTSwarm LLM-Debate MacNet Graph Early-Bird

Closest papers by meaning

• SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication (AAAI 2026 · similarity 0.91)
• MOC: Multi-Order Communication in LLM-based Multi-Agent Systems (ICML 2026 · similarity 0.89)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.89)
• Cost-Effective Communication: An Auction-based Method for Language Agent Interaction (AAAI 2026 · similarity 0.89)
• When Embedding-Based Defenses Fail: Rethinking Safety in LLM-Based Multi-Agent Systems (ICML 2026 · similarity 0.89)
• G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.88)