LLM-Driven Multi-Agent Systems

• LLM-driven MAS are systems where specialized LLM-powered agents dynamically coordinate to solve complex tasks using modular architectures.
• They employ diverse communication protocols, adaptive task graphs, and reinforcement learning to optimize collaboration and adaptability.
• These systems find applications in legal reasoning, engineering design, formal mathematics, and safety-critical autonomous operations.

LLM-driven multi-agent systems (MAS) constitute a rapidly maturing paradigm in which several (often specialized) LLM-powered agents autonomously interact, coordinate, and collaborate to solve complex tasks across domains such as legal reasoning, engineering design, code generation, formal mathematics, and robust control. These systems exploit LLMs’ generative and reasoning capacities while embedding them within agent architectures that support modularity, dynamic workflow composition, reinforcement learning, cross-agent critique, and domain adaptation. The resulting frameworks prominently extend the classical MAS tradition, offering flexible, scalable, and increasingly self-organizing systems that can match or surpass the problem-solving abilities of both single-agent LLMs and hand-engineered multi-agent pipelines.

1. Fundamental Models and System Architectures

LLM-driven MAS instantiate classic multi-agent principles—autonomous agenthood, communication, cooperation, and coordination—while leveraging LLMs as flexible, high-capacity reasoning modules. The formal models used to describe these systems range from classical Markov games with symbolic agent states and defined utility functions to fully generative, profile-driven agent tuples. Each LLM-agent is typified by a tuple

$A = \langle P, M, \mathcal{L}, T, C \rangle$

where the agent’s persona/profile ($P$), memory subsystem ($M$), pre-trained LLM backbone ($\mathcal{L}$), available external tools ($T$), and communication channels ($C$) are explicitly modularized (Bădică et al., 2 Sep 2025).

Modern LLM-MAS architectures exhibit diverse communication and workflow patterns:

• Centralized orchestration with a coordinator or manager agent (e.g., MAS-GPT, MAS$^2$ pipelines) (Ye et al., 5 Mar 2025, Wang et al., 29 Sep 2025).
• Decentralized peer-to-peer or hybrid topologies, often discovered or adapted at runtime (as in AMAS) (Leong et al., 2 Oct 2025).
• Dynamic or adaptive task graphs for asynchronous parallel execution (as in DynTaskMAS) (Yu et al., 10 Mar 2025).
• Parallelized planning-acting architectures, decoupling high-level task decomposition from atomic agent actions for real-time environments (Li et al., 5 Mar 2025).

Message-passing, shared and local memory, tool use, and agent-specific persona conditioning via prompt engineering are core operational principles. Inter-agent communication is often realized purely through natural-language dialogue, optionally augmented by role annotations, function tags, and structured JSON payloads (Bădică et al., 2 Sep 2025).

2. Methods for Collaboration, Coordination, and Workflow Optimization

LLM-MAS employ several methodologies to instantiate effective cooperative intelligence:

• Persona and expertise embedding: Agents are assigned distinct expert perspectives (engineering, legal, ethical, etc.), enabling specialized evaluation and multi-perspective critique (Mushtaq et al., 2 Jan 2025).
• Swarm-inspired consensus and negotiation: Utility-weighted voting, consensus updates $(x^{(t+1)} = \sum_i \alpha_i x_i^{(t)})$, and negotiation protocols enable the system to integrate diverse opinions and subspecialty judgments (Mushtaq et al., 2 Jan 2025).
• Dynamic task graph construction: Recursive decomposition of complex tasks into atomic subtasks aligned by precedence constraints, with dynamic adaptation as task requirements change (Yu et al., 10 Mar 2025).
• Parallel and interruptible planning-acting: Architectures decouple iterative high-level planning from rapid atomic execution, supporting preemption and real-time updates in dynamic environments (Li et al., 5 Mar 2025).
• Automated workflow/design generation: Systems such as MAS$^2$ and MAS-GPT employ generator-implementer-rectifier pipelines or code synthesis frameworks to produce, allocate, and refine MAS topologies and agent assignments adaptively, often using meta-agent routines and offline tree-based preference optimization (Wang et al., 29 Sep 2025, Ye et al., 5 Mar 2025).
• Heterogeneous LLM assignment: Agents may be powered by different LLMs selected to maximize function-domain performance, as in X-MAS (Ye et al., 22 May 2025).
• Adaptive topology selection: AMAS employs a lightweight LLM-based graph designer to select—per input—one among $K$ task-optimized agent graphs, learned by LoRA-based adaptation (Leong et al., 2 Oct 2025).

3. Learning and Optimization Strategies

Recent advances leverage both supervised and reinforcement learning, with the following methods prominent:

• Collaborative Reward Models (CRM): Fine-grained, stepwise reward feedback (e.g., correctness probability from a specialized PRM) enables dynamic agent selection and self-organization within task graphs, as in ReSo (Zhou et al., 4 Mar 2025).
• Multi-Agent Reinforcement Learning (MARL): Cooperative Dec-POMDP formulations and critic-free policy optimization (e.g., MHGPO, MAGRPO) supply scalable joint training of LLM agents, propagating system-level rewards backward to optimize inter-agent coordination (Chen et al., 3 Jun 2025, Liu et al., 6 Aug 2025).
• Cross-task experiential learning: Agents accumulate and retrieve high-reward stepwise experiences from prior tasks, supporting few-shot retrieval for current decision steps (as in MAEL), thus enabling generalization and rapid convergence (Li et al., 29 May 2025).
• Collaborative tree optimization: MAS$^2$ uses offline RL combined with value-scaled preference optimization to train generator, implementer, and rectifier meta-agents for robust self-generation and rectification of MAS (Wang et al., 29 Sep 2025).

Self-organization, stepwise critique, and reward-driven dynamic agent assignment increasingly characterize competitive large-scale LLM-MAS systems.

4. Application Domains and Benchmarking

LLM-MAS have been deployed in a range of specialized domains, including:

• Complex engineering and design review: Specialized persona agents enable rubric-aligned assessment and holistic critique in senior design projects (Mushtaq et al., 2 Jan 2025).
• Legal scenario synthesis and evaluation: MASER generates synthetic data and benchmarks for legal case interactions, integrating agent behavior alignment, distractor injection, and multi-stage evaluation (Yue et al., 8 Feb 2025).
• Mathematics autoformalization: MASA coordinates autoformalizer, critique, and refinement agents together with external theorem provers for high-fidelity translation of natural language mathematical statements to formal code (Zhang et al., 10 Oct 2025).
• Pest management and clinical decision support: PestMA demonstrates the value of editor-retriever-validator pipelines for high-accuracy pest management decisions, with similar principles proposed for healthcare (Shi et al., 14 Apr 2025).
• Safety-critical consensus control: Certified robustness against adversarial inputs and hallucinations is achieved via randomized smoothing and statistical certification in aerospace MAS (Hu et al., 5 Jul 2025).
• Research, code generation, and open-ended reasoning: MAS$^2$, MAS-GPT, ReSo, and other frameworks optimize over long-horizon, error-prone tasks where adaptation and resilience are critical (Wang et al., 29 Sep 2025, Ye et al., 5 Mar 2025, Zhou et al., 4 Mar 2025).

Benchmarking is advanced through MAS-specific datasets and evaluation protocols (e.g., MILE, SynthLaw, MAEL benchmarks), assessing accuracy, interactivity, logicality, efficiency, and cross-backbone generalization (Yue et al., 8 Feb 2025, Li et al., 29 May 2025, Ye et al., 22 May 2025). Notably, heterogeneous agent assignment and adaptive topology deliver marked performance gains over homogeneous or static graph baselines (Ye et al., 22 May 2025, Leong et al., 2 Oct 2025).

5. Robustness, Security, and Scalability

As LLM-MAS mature, robustness and security concerns demand rigorous treatment:

• Randomized smoothing and robustness certification: These techniques guarantee agent-level and system-level invariance to bounded adversarial perturbations and LLM hallucinations, with certifiable consensus error reductions (89% reduction in aerospace MAS) (Hu et al., 5 Jul 2025).
• Web Fraud Attacks: MAS architectures are vulnerable to stealthy URL-based attacks (e.g., homoglyphs, parameter manipulation) with up to ~94% success rates against state-of-the-art defenses, exposing the need for dedicated semantic link-validation modules combining Unicode normalization and token analysis (Kong et al., 1 Sep 2025).
• Noise and memory management: Analytical probabilistic models of shared vs. separate context management (as in the Response Consistency Index) reveal trade-offs in consistency, speed, and noise resilience, providing design guidelines for scalable MAS under memory and communication constraints (Helmi, 9 Apr 2025).

Empirical studies underline that system performance and reliability hinge on both dynamic agent adaptation and rigorous, domain-specific safety mechanisms.

6. Open Challenges and Future Directions

LLM-MAS research now faces several open challenges and lines of attack:

• Scalability: Dialogue- and language-heavy protocols may not scale to thousands of interacting agents; asynchronous scheduling and adaptive context management offer partial mitigation (Yu et al., 10 Mar 2025, Helmi, 9 Apr 2025).
• Interpretability and verification: Bridging the gap between black-box generative reasoning and verifiable symbolic plans remains open, motivating hybrid BDI–LLM approaches and integration of external formal proof/checking systems (Zhang et al., 10 Oct 2025, Bădică et al., 2 Sep 2025).
• Standardization and interoperability: Absence of common agent, message, and workflow protocols hinders reproducibility and integration; new benchmarks and code representations (e.g., as executable Python classes) are emerging (Ye et al., 5 Mar 2025, Wang et al., 29 Sep 2025).
• Security: Evolving adversarial vectors—especially web-link and prompt-based attacks—demand integrated, layered defenses drawing on domain knowledge, tokenization, and behavioral validation (Kong et al., 1 Sep 2025).
• Generalization and meta-adaptation: Meta-agent architectures (as in MAS$^2$), experiential retrieval (MAEL), and self-organizing agent selection (ReSo) point the way to continually self-evolving, cross-backbone generalizable LLM-MAS (Li et al., 29 May 2025, Wang et al., 29 Sep 2025, Zhou et al., 4 Mar 2025).
• Human–AI teaming: Open questions remain on optimal interfaces, division of labor, and negotiation between human and LLM-based agents.

In sum, LLM-driven MAS synthesize and extend core MAS theory with generative, adaptive, and collaborative intelligence. Through modular agent construction, dynamic workflow orchestration, reinforcement- and experience-driven optimization, and domain-tailored safety engineering, they are rapidly establishing themselves as a fundamental building block for next-generation intelligent systems across scientific, engineering, legal, and safety-critical domains (Bădică et al., 2 Sep 2025, Li et al., 29 May 2025, Hu et al., 5 Jul 2025, Wang et al., 29 Sep 2025, Ye et al., 22 May 2025, Ye et al., 5 Mar 2025, Leong et al., 2 Oct 2025, Chen et al., 3 Jun 2025, Liu et al., 6 Aug 2025, Zhou et al., 4 Mar 2025, Helmi, 9 Apr 2025).