Coordinating from Memory: Graph-Structured Experience Reuse for Multi-Agent Adaptation in Dynamic Manufacturing

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

MANTA, a framework for Multi-Agent Network Topology Adaptation that enables communication structures to self-evolve at inference time, is introduced and shows that inference-time self-improvement can extend to the architecture of collaboration itself.

STAGE: Spatio-Temporal Aggregation via Graph Embedding for Multi-Agent Reinforcement Learning in Industrial Optimization

Industrial multi-agent coordination requires distributed subsystems to collaborate under heterogeneous relationship structures whose relative importance shifts across operational contexts—physical constraints dominate startup while operational hierarchies govern steady-state. Existing multi-agent reinforcement learning approaches either ignore these structural distinctions or aggregate them uniformly, limiting adaptive coordination capabilities. This paper presents STAGE (Spatio-Temporal Aggregation via Graph Embedding), integrating multi-layer graph processing with spatio-temporal learning for context-dependent coordination. The architecture processes distinct relationship types through dedicated attention mechanisms with learned adaptive fusion, enabling coordination emphasis to adjust dynamically across operational phases. Spatio-temporal integration couples multi-layer spatial structures with multi-scale temporal dynamics through attention-based fusion mechanisms, while graph-guided hypernetworks generate mixing weights that preserve the Individual-Global-Max property essential for decentralized execution. Comprehensive evaluation on steam power plant coordination demonstrates that STAGE significantly outperforms existing multi-agent and optimization methods in both learning efficiency and asymptotic performance, while providing interpretable coordination mechanisms and maintaining the monotonicity property critical for decentralized industrial deployment. Note to Practitioners—Industrial subsystems interact through multiple relationship types whose importance varies across operational contexts. We propose a framework that separately processes these relationships and learns to adjust their emphasis adaptively. Graph-guided coordination ensures individual actions optimize system-wide performance. Evaluation on power plant configuration demonstrates superior results over conventional optimization and learning methods. The approach learns from operational data while respecting the simulator’s physical feasibility constraints. We believe this framework has strong potential for chemical processing, manufacturing, and energy management systems.

Adaptive AI Agent Migration via Generative Diffusion-based Reinforcement Learning in Edge Intelligence Systems

Deploying Large Language Model (LLM)-based AI agents at the network edge enables real-time task handling in 6G networks. However, resource heterogeneity and user mobility necessitate agent migration to maintain Quality of Service (QoS). Unlike stateless containers, AI agents encapsulate invocation histories, planning contexts, and memory stores, making their migration considerably more complex. To address this issue, we propose AMD, an adaptive AI Agent Migration framework that integrates conditional Diffusion models with reinforcement learning to jointly optimize latency and resource utilization under dynamic edge environments. Specifically, we introduce a diffusion model as a generative prior to produce high-quality global deployment plans, effectively avoiding the local-optima problem common in conventional reinforcement learning. A reinforcement learning-based module then performs online policy refinement to accommodate real-time environmental variations. We implement AMD on a distributed system built upon AgentScope and validate it across geographically distributed edge servers. Experimental results demonstrate that AMD reduces the average task latency by 3.9%–27.6% and improves resource utilization by up to 53.3% compared to baseline strategies.

Reward-Driven LLM Agent Workflows: Synthesizing POMDP Routing and Self-Correction for Autonomous Decision-Making

This paper addresses key technical challenges in current large language model (LLM) agent applications, including long-horizon planning, sparse reward attribution, and dynamic environmental interaction, by designing and optimizing an intelligent agent workflow. The proposed architecture is based on the synthesis of core AI paradigms: Visual, Language, Generative, Graph, Multimodal, Reinforcement, and Agent Intelligence. Unlike conventional baseline models that rely on static prompting and lack robust perception-action loops, our approach introduces a Partially Observable Markov Decision Process (POMDP) routing mechanism. This mechanism is augmented with an internal, self-correcting reward model that evaluates decision trajectories before execution. By integrating multimodal inputs and advanced reinforcement learning principles (such as proximal policy optimization and value function approximation), the agent maintains long-term structural memory and dynamically adapts its reasoning pathways to mitigate error accumulation. Empirical experiments on the ALFWorld embodied simulation environment and the WebShop online navigation benchmark demonstrate a 24.5% absolute improvement in task success rate and trajectory efficiency over mainstream baselines like the standard ReAct framework. Comprehensive ablation studies confirm the significant contribution of the reward-driven critique module in suppressing hallucination rates. This research bridges theoretical foundations of reinforcement learning and graph-based memory with autonomous agent workflows. Ultimately, the resulting architecture offers a practical, scalable reference framework for developing artificial intelligence technologies in complex, multi-step autonomous systems. Code is available at https://github.com/01Amez/RLAW_Implementation.

Multiagent Dynamic Task Allocation Based on Graph Neural Reinforcement Learning Algorithm

Abstract Multi-agent dynamic task allocation (MADTA) for UAV swarm and autonomous systems remains a formidable challenge in highly uncertain and stochastic environments, where conventional reinforcement learning methods struggle with variable input dimensions and coordination conflicts. This paper proposes a spatiotemporal topology-aware graph reinforcement learning (STA-GRL) framework to address these limitations. By modeling the environment as a dynamic bipartite graph, the framework integrates a spatiotemporal gated graph attention (STGGA) module that employs a temporal gating mechanism to dynamically prioritize tasks with rapidly decaying deadlines. A topology-aware critic is further designed to penalize spatial conflicts among agents via an enhanced adjacency matrix. Extensive simulations demonstrate that STA-GRL significantly surpasses state-of-the-art baselines. In the primary evaluation scenario with 30 agents and an intermediate task arrival rate (λ = 0.6), STA-GRL achieves a task completion rate of 86.8% and an average response time of 18.4 seconds, while reducing the average conflict rate to just 2.1%. Moreover, ablation studies confirm the critical contribution of each architectural component, with the temporal gate improving the completion rate by 7.3% and the topology-aware critic reducing conflicts by 6.4%.

Bidirectional job-shop scheduling with graph-theoretic features: a graph attention network and proximal policy optimization approach