MasFACT: Continual Multi-Agent Topology Learning via Geometry-Aware Posterior Transfer

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

MasFACT is a novel geometry-aware posterior transfer framework designed to address "topology forgetting" in Large Language Model (LLM)-based multi-agent systems (MAS). This issue arises when MAS adapt to new tasks, causing their communication topology generators to lose previously effective collaboration patterns. MasFACT preserves historical collaboration knowledge as transferable topology priors, aligning them across task-specific agent spaces using Fused Gromov-Wasserstein optimal transport. It then performs PAC-Bayes-guided conservative posterior adaptation to balance task plasticity with structural stability. Experiments across 12 continual scenarios, spanning class-, domain-, and task-level settings, demonstrate that MasFACT improves average accuracy by 9.3% and reduces average forgetting by 69.4% compared to strong baselines, and integrates seamlessly with various MAS topology generators like GDesigner, AgentDropout, and ARGDesigner, using Llama-3.1-8B-Instruct.

Key takeaway

For AI Scientists and Machine Learning Engineers designing multi-agent systems for evolving task streams, you should consider integrating MasFACT to mitigate "topology forgetting." This framework allows your MAS to retain and reuse effective collaboration patterns from prior tasks, improving average accuracy by 9.3% and reducing forgetting by 69.4%. Implement its geometry-aware prior transfer and PAC-Bayes-guided adaptation to ensure structural stability while maintaining task-specific plasticity.

Key insights

MasFACT prevents "topology forgetting" in multi-agent systems by reusing historical collaboration patterns as geometry-aware priors.

Principles

• Topology forgetting is a distinct structural failure mode.
• Jointly align relational structure and agent semantics.
• Balance task plasticity with structural stability.

Method

MasFACT constructs factorized prior atoms, retrieves and aligns the most transferable prior via Fused Gromov-Wasserstein optimal transport, then learns a sparse topology posterior around it with PAC-Bayes control.

In practice

• Integrate MasFACT with existing MAS topology optimizers.
• Use FGW to align agent attributes and graph structures.
• Apply sparse residual edits for task-specific adaptation.

Topics

• Multi-Agent Systems
• Continual Learning
• Topology Forgetting
• Fused Gromov-Wasserstein
• PAC-Bayes
• LLM Communication Topology

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Bounding the Predictive Space: How Topological AI Solves Catastrophic Forgetting Through… — Topological AI uses prime-indexed embedding anchors to prevent catastrophic forgetting in LLMs while retaining cognitive flexibility.
• MPCS: Neuroplastic Continual Learning via Multi-Component Plasticity and Topology-Aware EWC — Integrating multiple plasticity mechanisms and using Pareto analysis can optimize continual learning systems.
• From Order to Distribution: A Spectral Characterization of Forgetting in Continual Learning — Task distribution geometry, not just task order, fundamentally governs forgetting rates in continual learning.
• Teaching AI to Remember: What I Learned Building a Continual Learning Framework for Drug Discovery — Continual learning models can avoid catastrophic forgetting by integrating new knowledge with awareness of existing representations.
• Continual Learning with Elastic Regularization and Synthetic Replay for Federated MLLM Fine-Tuning — FedCMM prevents catastrophic forgetting in federated MLLM fine-tuning by embedding multi-level continual learning safeguards.
• The Stable Recovery Manifold: Geometric Principles Governing Recoverability in Continual Learning — Catastrophic forgetting stems from accessibility and manifold-alignment issues, not knowledge destruction, as forgotten information remains compactly decodable.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.