ASMAD: Adaptive Sparse Communication Topology Multi-Agent Debate Framework with Opinion Dynamics

ACL ARR 2025 May Submission924 Authors

Abstract: Large language models (LLMs) still face challenges in complex reasoning within multi-agent debate (MAD) systems due to high computational costs in fully-connected structures. While existing methods use static sparse topologies to reduce computation, they neglect semantic relationships and dynamic opinion evolution. To solve this challenge, we propose ASMAD, an adaptive sparse topology framework that synergizes sociophysical opinion dynamics with LLMs through two innovations: (1) probabilistic semantic-guided attention gates for dynamic opinion visibility control; (2) a hybrid paradigm combining adaptive trust-boundary regulation and opinion synchronization. Experiments show ASMAD reduces token costs to around 1/3 across GSM8K and MMLU benchmarks while maintaining competitive accuracy with 4-bit quantized 7-9B size models.

Paper Type: Long

Research Area: Efficient/Low-Resource Methods for NLP

Research Area Keywords: LLM Efficiency, NLP in resource-constrained settings,

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings

Languages Studied: English

Keywords: LLM Efficiency, NLP in resource-constrained settings

Submission Number: 924