To Be or Not to Be: Bayesian Survival Pruning for Multi-Agent System

ACL ARR 2025 May Submission4317 Authors

Abstract: To address the challenges of communication topology redundancy and Newcomb's paradox in open-world multi-agent systems, we propose Agent Bayesian Out(ABO), an uncertainty-aware framework. By modeling edge weights in communication graphs as probabilistic random variables, ABO innovatively integrates three key mechanisms: Bayesian edge weight representation, Gaussian process priors, and Markov Chain Monte Carlo-based graph sampling. This integration establishes a dynamic uncertainty propagation model, overcoming the limitations of conventional methods in stochastic modeling and open-scenario adaptation. Experiments demonstrate that ABO effectively identifies low-contribution nodes and redundant connections, achieving an average accuracy improvement of 1.8%-2.59% on benchmarks. Notably, it exhibits enhanced transfer capability in open-domain QA tasks. Ablation studies confirm the synergistic effects of components through a Bayesian-uncertainty sampling mechanism, while its zero additional inference overhead provides a novel paradigm for distributed agent collaboration. The proposed framework outperforms existing state-of-the-art (SOTA) methods across multiple metrics and achieves cutting-edge performance.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Multi-Agent System, Uncertainty Modeling, Directed Graphical Model

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English, Chinese

Submission Number: 4317