2026· Annual Meeting of the Association for Computational Linguistics· pp. 27553-27574· 0 citations· 52 references
Computer Science
TL;DR
Hetero-Designer is proposed, a novel pipeline that efficiently encodes intricate dependencies among queries, LLMs and roles through a novel Binary-Star Transformer and constructs Hetero-MAS in an autoregressive graph generation process.
Abstract
LLM-based Multi-agent systems (MAS) have shown strong capabilities across a wide range of domains. Their success largely hinges on the collaboration topology design, which has emerged as a central research focus in the automated MAS design. However, existing approaches are fundamentally constrained by their reliance on homogeneous LLMs, which significantly limits overall system intelligence. In response to this limitation, we for the first time propose the concept of Automated Design of Heterogeneous-LLMs-based MAS (ADHM) . ADHM sheds light on a promising avenue for advancing collective intelligence, which focuses on the automated design of cost-effective MAS composed of diverse LLMs and roles to suit various queries. Toward this challenging goal, we propose Hetero-Designer , a novel pipeline that efficiently encodes intricate dependencies among queries, LLMs and roles through a novel Binary-Star Transformer and constructs Hetero-MAS in an autoregressive graph generation process. Extensive experiments demonstrate that Hetero-Designer is:
This paper distill four design principles for scalable MAS architectures from a structured analysis of prior work: simplicity, elastic feedback, sequential workflows with optional loops, and summary-based communication, and evaluates four configurations of increasing complexity on a standardized benchmark of terminal-based system engineering tasks.
Linus Sander, Feng Pan, Vahid Zolfaghari et al.· arXiv.org· 0 citations
This work presents OptiMAS, a task-agnostic agentic optimizer that leverages textual interaction trajectories and task feedback as loss signals for end-to-end MAS evolution and sustains performance improvement over extended optimization horizons.
Yuxin Cheng, Chang Liu, Hanxin Yu et al.· 0 citations
CEDAR is an autonomous method that uses Large Language Model agents to discover complex systems satisfying user-specified behavioral goals, enabling goal-directed discovery of complex system behaviors while preserving solution diversity, and its LLM-based interpretability reveals how structural changes drive emergent behavior.
Ambient Assisted Living (AAL) environments aim to enable older people to remain active and lead an autonomous and independent life for as long as possible. Among the technologies that can be incorporated into these settings, socially assistive robots (SAR) seek to establish a more natural and intuitive means of interaction with people, whilst helping them to carry out everyday tasks. One of the main challenges facing the design of these robots is how to enable them to undertake more complex tasks. Recent advances in Large Language Models (LLMs) have opened new avenues for flexible robot deliberation, yet their integration into real-time robotic systems remains challenging due to latency constraints, reasoning reliability, and the complexity of coordinating multi-step tasks. This paper proposes a hierarchical multi-agent architecture for robot deliberation that addresses these challenges by combining LLM-based planning with structured execution mechanisms within the ROS 2 ecosystem. The proposed architecture employs a supervisor agent that decomposes high-level natural language instructions into prioritised subtasks, enabling a priority-driven execution model that dynamically adapts to task relevance, temporal constraints, and environmental feedback. Subtasks are delegated to a set of Single-Purpose Agents (SPAs), orchestrated via LangGraph state machines and coordinated through a priority-aware scheduling mechanism. A key design principle is the use of Behaviour Trees (BTs) as high-level callable tools through the Model Context Protocol (MCP), encapsulating closed-loop control strategies while enabling preemptive and priority-consistent execution. This reduces the number of LLM inference steps required per task and improves robustness under dynamic conditions. A further contribution concerns the deployment of fine-tuned, lightweight LLMs—on the order of 0.6 billion parameters—specifically adapted for both the supervisor and the individual SPA roles through parameter-efficient low-rank adaptation (LoRA). These models are trained on role-specific tool-calling datasets to specialise in constrained reasoning patterns and task-specific decision-making, enabling efficient, low-latency inference directly on edge hardware. The combination of fine-tuning and hierarchical priority control enhances both the determinism and responsiveness of the system while mitigating error propagation across agent interactions. The paper presents the full software architecture, a formal characterisation of the system as a priority-aware hierarchical policy over a graph of agent workflows, and an experimental evaluation in an Ambient Assisted Living scenario assessing task success rate, inference efficiency, responsiveness under competing priorities, and overall user experience. Because SPA execution is decoupled from the supervisor’s own reasoning loop, the architecture is designed to keep accepting, processing, and queuing new user queries while previously dispatched SPAs are still executing their tasks.
Alberto J. Tudela, Óscar Pons, José Galeas et al.· Applied Sciences· 0 citations
Large Language Models (LLMs) have demonstrated remarkable capabilities as powerful components in agentic systems, enabling sophisticated reasoning and complex task execution. However, current approaches to manually designing and optimizing agentic systems heavily rely on manual effort, limiting their adaptability and scalability. Recent work has explored the automated optimization of workflow designs. However, these approaches often overlook the crucial role of model capabilities and focus on single performance metrics, failing to address real-world deployment constraints. In this paper, we present AgentFactory, a framework that jointly optimizes both foundation models and workflow structures in agentic systems while considering multiple objectives including performance, cost, and efficiency. AgentFactory leverages advanced LLMs as optimizers to navigate the vast search space of possible configurations, employing a three-stage optimization pipeline to automatically discover effective combinations of fine-tuned models and optimized workflows. Through an iterative optimization process, our framework systematically explores and evaluates different agentic system designs, adapting to task-specific requirements while maintaining operational efficiency. We evaluate AgentFactory across eight benchmarks spanning five domains, including general reasoning, coding, mathematics, medicine, and finance. Our experiments demonstrate that AgentFactory consistently outperforms both manually designed methods and existing automated approaches, achieving an average improvement of 9.1% across all benchmarks, with particularly significant gains in domain-specific tasks (19.6% on MedQA and 18.7% on FinEval). These results establish AgentFactory as a promising approach for developing more capable and efficient agentic systems through automated optimization.
En-Ci Zhang, Hao-Fen Wang, Yuesheng Zhu et al.· Pacific Rim International Co...· 0 citations
LLM powered multi-agent systems (MAS) have emerged as a promising paradigm for complex tasks. However, their advantages over single-agent systems (SAS) remain unclear, with performance varying inconsistently across settings. Here, we provide an information bottleneck perspective on elucidating the differences between MAS and SAS. Specifically, our key observation is that a SAS accumulates its full reasoning trace in one shared context, while a MAS uses isolated local contexts connected by bounded relay messages. We show that, under infinite relay bandwidth, any SAS can be simulated by a MAS that transmits the full upstream context. Thus, the nontrivial advantage of MAS arises under bounded relays, where compression introduces a fundamental trade-off: reducing redundant context can improve efficiency, but may also incur loss of task-relevant information. We formalize this trade-off as an information bottleneck controlled by an effective parameter $\beta$, which captures how the balance shifts with model capability, and shows that MAS gains arise when context reduction outweighs relay information loss. We conduct 18 controlled experiments across five benchmarks and three model scales to validate our theoretical studies. We observe that MAS consistently helps when relays are near-sufficient, especially for weaker models. In contrast, MAS gains shrink or reverse when relays incur information loss, especially for stronger models that can already extract useful information from redundant context and thus gain little from compression. Our study shows that multi-agent design is fundamentally an information-bottleneck optimization problem. This perspective explains when bounded inter-agent communication helps or hurts.
