PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems

Shihao Qi, Jie Ma, Rui Xing, Wei Guo, Xiao Huang, Zhitao Gao, Jianhao Deng, Jun Liu, Lingling Zhang, Bifan Wei, Boqian Yang, Pinghui Wang, Jianwen Sun, Jing Tao, Yaqiang Wu, Hui Liu, Yu Yao, Tongliang Liu

arXiv:2605.14892v2cs.AI

TL;DR

LLM-based multi-agent systems face coordination and failure-diagnosis challenges as errors propagate through interactions. This survey addresses the gap with the LIFE framework, synthesizing four stages and their causal dependencies to provide a cross-stage roadmap for self-improving systems.

  • Problem

    Sustained multi-agent coordination remains difficult because interaction-driven anomalies can propagate and amplify across agents and execution steps.

  • Method

    The survey organizes multi-agent systems around the causally linked LIFE progression and provides systematic taxonomies, comparative analyses, and cross-stage dependency characterizations.

  • Results

    The survey unifies four stages—capability foundations, collaboration, failure attribution, and self-evolution—and traces the causal dependencies linking them.

  • Takeaways & Limitations

    The LIFE framework makes explicit how collaborative structures shape observable failures and supports analysis of collective capabilities that can exceed individual agents.

  • Takeaways & Limitations

    Scalability remains a major bottleneck because adding agents does not guarantee better performance and can increase reliability and cost challenges.

Abstract

from arXiv · show

LLM-based autonomous agents have demonstrated strong capabilities in reasoning, planning, and tool use, yet remain limited when tasks require sustained coordination across roles, tools, and environments. Multi-agent systems address this through structured collaboration among specialized agents, but tighter coordination also amplifies a less explored risk: errors can propagate across agents and interaction rounds, producing failures that are difficult to diagnose and rarely translate into structural self-improvement. Existing surveys cover individual agent capabilities, multi-agent collaboration, or agent self-evolution separately, leaving the causal dependencies among them unexamined. This survey provides a unified review organized around four causally linked stages, which we term the LIFE progression: Lay the capability foundation, Integrate agents through collaboration, Find faults through attribution, and Evolve through autonomous self-improvement. For each stage, we provide systematic taxonomies and formally characterize the dependencies between adjacent stages, revealing how each stage both depends on and constrains the next. Beyond synthesizing existing work, we identify open challenges at stage boundaries and propose a cross-stage research agenda for closed-loop multi-agent systems capable of continuously diagnosing failures, reorganizing structures, and refining agent behaviors, extending current coordination frameworks toward more self-organizing forms of collective intelligence. By bridging these previously fragmented research threads, this survey aims to offer both a systematic reference and a conceptual roadmap toward autonomous, self-improving multi-agent intelligence.

5 Multi-Agent System Self-Evolution

Section 5 develops multi-agent system self-evolution from attribution through formalization, taxonomy, evolutionary dynamics, evaluation, and discussion. It distinguishes agentic, systemic, and meta self-evolution within its taxonomy.

  • From Attribution to Evolution; Formal Definition of MAS Self-Evolution: The section begins by connecting fault attribution to multi-agent system self-evolution and formally defining the concept.
  • MAS Self-Evolution Taxonomy: Its taxonomy separates MAS self-evolution into agentic, systemic, and meta levels.
  • Analyzing Evolutionary Dynamics; Evaluation: The section analyzes evolutionary dynamics and then evaluates self-evolution approaches.
  • Discussion: It concludes with a discussion of multi-agent system self-evolution.

6 Conclusion 64 · 1 Introduction

LLM-based multi-agent systems extend individual agent capabilities through collaboration, but inter-agent dependencies make failures harder to diagnose and improve. The survey unifies capabilities, collaboration, failure attribution, and self-evolution into the causally linked LIFE progression and proposes an attribution-driven closed loop.

  • 1 Introduction: LLMs have advanced from fluent text generation to complex reasoning, long-horizon planning, and interaction with external environments.
  • 1 Introduction: Extended real-world tasks remain difficult because agents struggle with coherent behavior, changing conditions, and multi-step actions across tools and environments.
  • 1 Introduction: Multi-agent collaboration addresses single-agent limitations by decomposing complex tasks into specialized roles coordinated through structured communication and orchestration.
  • 1 Introduction: Tighter interdependence makes collective behavior increasingly difficult to predict, diagnose, and improve when localized errors propagate into cascading failures.
  • 1 Introduction: Reliable diagnosis and actionable self-correction are mutually necessary for closed-loop multi-agent systems that improve autonomously.
  • 1 Introduction: Existing surveys separately examine individual capabilities, multi-agent collaboration, or self-evolution, while largely omitting multi-agent failure attribution and the complete operational lifecycle.
  • 1 Introduction: The survey presents LIFE as a unified lens linking individual intelligence, collaboration, failure attribution, and self-evolution across four causally linked stages.
  • 1 Introduction: LIFE supplies stage-wise taxonomies, an attribution–evolution feedback loop, boundary challenges, a cross-stage agenda, and curated research resources.

2 Individual Intelligence · 2.1 From LLM to LLM-based Agent: An Architectural Overview · 2.2 Reasoning

The section frames individual intelligence as the foundation of autonomous agents, which extend LLMs with iterative perception, reasoning, action, and reflection. It then organizes reasoning enhancement around input expansion, process search and verification, and output regulation.

  • 2.1 From LLM to LLM-based Agent: An Architectural Overview: LLM-based agents address LLM limitations in complex tasks by perceiving states, formulating goals, using external tools, and adapting through feedback.They operate through an iterative loop of observation, reasoning, action, and reflection.
  • 2.1 From LLM to LLM-based Agent: An Architectural Overview: Individual agent intelligence comprises four interacting capabilities: reasoning, memory, planning, and tool use.The resulting policy is factorized into specialized modules, enabling structured and interpretable decision-making.
  • 2.1 From LLM to LLM-based Agent: An Architectural Overview: An agent is modeled as a sequential decision-making system that receives observations, maintains memory, reasons over context, plans actions, executes tools, and updates memory.Its tuple is a = (X, U, O, M, R, P, T), with R, P, and T denoting reasoning, planning, and tool execution.
  • 2.2 Reasoning: Reasoning is the cognitive foundation of agent decision-making, and enhancement methods target information access, reasoning trajectories, and generated-output reliability.These dimensions correspond to input-stage enhancement, reasoning-process enhancement, and output-stage regulation.
  • 2.2.1 Input-Stage Enhancement: Input-stage enhancement supplements parametric knowledge with external information through knowledge augmentation and modality augmentation.Examples include retrieval, adaptive critique, multimodal chain-of-thought, visual artifacts, and step-by-step visual manipulation.
  • 2.2.2 Reasoning-Process Enhancement: Reasoning-process enhancement expands candidate search through chains, trees, and graphs, then applies path verification to select reliable solutions.DeepSeek-R1 enables autonomous extended reasoning through reinforcement learning, while process reward models score individual reasoning steps.
  • 2.2.3 Output-Stage Regulation: Output-stage regulation addresses hallucination through reliability assessment and response-behavior regulation, including verification, decoding adjustments, alignment, refusal, abstention, and tool-based checking.Reliability signals may come from external evidence, agreement across samples, model internals, or calibrated confidence.

2.3 Memory

Agent memory converts interaction experience into organized, retrievable information that supports better future actions. The memory lifecycle comprises formation, maintenance, and retrieval/utilization, with design choices shaping downstream reasoning, planning, and skill reuse.

  • Memory lifecycle: Memory forms a closed loop in which interaction produces experience, experience is consolidated into memory, and retrieved memory informs later actions.This memory scope concerns agent-level experiential memory rather than parametric knowledge or static external knowledge bases.
  • Memory formation: Memory formation transforms interaction streams into storable entries through direct recording, abstractive distillation, or typed routing.Typed routing extends abstraction by assigning entries to episodic, semantic, or procedural stores.
  • Memory formation: Topically coherent segments with compression-based denoising outperform both turn-level and session-level alternatives in memory quality.This finding indicates that segmentation granularity critically affects abstractive memory construction.
  • Memory formation: Memory type determines downstream capabilities: episodic traces support reasoning-time self-correction, semantic knowledge informs experience-driven planning, and procedural entries enable cross-task skill reuse.Collapsing typed stores into a single store substantially degrades retrieval quality.
  • Memory maintenance: Memory maintenance manages unbounded memory growth under limited attention through storage structures and dynamic mechanisms that update, compress, or discard memories as relevance changes.Flat append-only storage limits scalability, while maintenance spans passive decay to active policy learning.

2.4 Planning

Planning converts high-level goals into executable action sequences while accounting for dependencies, feedback, and failures over extended horizons. Existing methods span decomposition-based planning, which structures tasks around execution, and search-based planning, which evaluates alternative paths.

  • Planning converts high-level goals into structured executable action sequences while accounting for sub-task dependencies, environmental feedback, and potential failures over extended horizons.
  • Two paradigms organize planning methods: decomposition-based planning structures goals into sub-tasks, whereas search-based planning evaluates alternative plans through systematic evaluation and backtracking.
  • Proactive Decomposition: Proactive decomposition generates complete plans before execution, including dependency graphs, parallelizable function-call DAGs, grounded plans, and explicit high-level plans for long-horizon web tasks.Plan-and-Act demonstrates that explicit proactive planning significantly outperforms reactive approaches on complex web benchmarks.
  • Progressive Decomposition: Progressive decomposition interleaves planning with execution, refining plans from observations through reasoning-action loops, reflection on failures, recursive sub-task decomposition, and dynamic goal-tree refinement.Voyager extends this paradigm with incremental planning and a growing verified skill library for continual capability expansion.
  • Search-Based Planning: Search-based methods explore multiple candidate paths or plans, using mechanisms such as tree search, MCTS, A*, retrieval-based reranking, voting, scoring, and verification.

2.5 Tool Use

Tool use expands an agent’s action space by connecting language reasoning to external execution, while requiring structured calls, result processing, and integration into the reasoning flow. The section organizes tool use around capability acquisition, invocation control flow, and generalization to unseen tools and environments.

  • Overview: Tool use extends agents beyond text generation by invoking external functions, APIs, code interpreters, web services, and other systems.Unlike internal reasoning and planning, tool use crosses from language into execution through structured calls and external results.
  • Overview: The tool-use lifecycle comprises capability acquisition, invocation, and generalization across unseen tools and environments.These dimensions cover how models learn tool use, how tools are called during execution, and how abilities transfer beyond training conditions.
  • Tool Capability Acquisition: Capability acquisition requires recognizing when tools are needed, selecting appropriate tools, generating valid parameters, and incorporating returned results.Approaches include trajectory-based learning from demonstrations and protocol-based alignment with tool interface descriptions.
  • Tool Invocation: Tool invocation progresses from one-shot calling to closed-loop interaction and workflow-based orchestration as coordination complexity increases.Closed loops use tool outputs to inform later decisions, while workflows explicitly govern invocation order and data flow for multi-tool goals.
  • Tool Generalization: Tool generalization addresses unseen-tool adaptation and relevant-tool discovery from large, evolving collections.Documentation comprehension, retrieved specifications, and self-verification support transfer beyond memorized call patterns; ToolLLM shows zero-shot generalization on out-of-distribution APIBench.

2.6 Evaluation · 2.7 Discussion

Evaluation has shifted toward integrated, trajectory- and subgoal-level assessment, but important gaps remain in cross-capability and resource-aware benchmarking. The discussion identifies unresolved coupling across modules, reasoning and reliability, memory design, and continuous adaptation.

  • 2.6 Evaluation: Benchmarks increasingly jointly assess reasoning, planning, and tool use in unified task settings rather than isolated skills.Evaluation criteria also increasingly include process-level signals such as action trajectories, subgoal completion, and execution traces.
  • 2.6 Evaluation: Comprehensive suites span diverse environments and multi-turn interactions, with newer benchmarks measuring partial task completion and trajectory-level behavior.AgentBench evaluates eight environments, while AgentBoard introduces subgoal-based progress metrics; GAIA and MINT combine web interaction with tool use.
  • 2.6 Evaluation: Specialized benchmarks now cover web, operating systems, software engineering, tool use, memory, scientific research, enterprise applications, and safety-critical tasks.Examples include WebArena, OSWorld, SWE-bench, LoCoMo, MLE-bench, WorkArena++, Cybench, and AgentHarm.
  • 2.6 Evaluation: Current evaluation rarely tests cross-capability effects or incorporates token consumption, latency, and cost, motivating integrated, resource-aware protocols.The capability bottlenecks exposed by single-agent benchmarks also motivate multi-agent collaboration strategies.
  • 2.7 Discussion: Modular optimization dominates despite interdependent reasoning, memory, planning, and tool use, leaving capability mismatches invisible in benchmarks targeting only one or two capabilities.These mismatches often become the actual bottleneck in end-to-end deployment.
  • 2.7 Discussion: Reasoning enhancement and hallucination suppression remain methodologically separated: broader search increases error exposure, process rewards scale poorly, and reliability assessment is mostly post-hoc.Dynamic inference-time steering based on step-level confidence remains unresolved.
  • 2.7 Discussion: Memory research lacks systematic analysis of interactions among formation, storage, maintenance, and retrieval, while storage choices often ignore downstream compatibility.Graph-based storage has high maintenance cost, and utility-driven maintenance requires extensive interaction to converge.
  • 2.7 Discussion: Capability evaluations remain static snapshots, treating tool generalization and memory evolution as one-time capacities rather than continuous adaptation required in deployment.The reviewed methods assume frozen strategies and known tool sets, whereas real deployment demands ongoing adaptation.

3 Multi-Agent Collaboration

Multi-agent systems extend problem-solving beyond individual agents by integrating agents with distinct capabilities. The section formalizes collaboration as a structured, goal-oriented paradigm defined by Role, Communication, Orchestration, and Interaction.

  • Multi-agent systems use cooperation among agents with distinct capabilities to tackle tasks beyond any individual agent’s ability.
  • The collaboration framework decomposes multi-agent systems into four components: Role, Communication, Orchestration, and Interaction.Together, these components address who acts, how agents communicate, how they are connected, and how they collaborate during execution.
  • Role: Role specialization assigns agents responsibilities such as planning, reasoning, or execution through a role allocation function.The formulation represents each agent’s role as z_i = ρ(a_i), with roles drawn from the set Z.
  • Communication: Communication governs how agents generate, structure, and transmit messages using local interaction histories and accessible shared context.Messages follow protocol C and are generated from information H_i,t available to agent a_i at time t.
  • Orchestration and Interaction: Orchestration defines agent connectivity and message-passing pathways, while Interaction specifies coordinated execution under a collaboration policy and environment.Interaction describes agents acting, exchanging messages, and coordinating as a whole; evaluating resulting trajectories provides a unified basis for comparing designs.

3.1 Role

Role in multi-agent collaboration comprises role capability—what agents can do and access—and role allocation—how responsibilities are assigned among them. Role capabilities may be homogeneous or heterogeneous, while allocation may be static or dynamic, shaping coordination, specialization, and adaptability.

  • Role dimensions: Role decomposes into Role Capability and Role Allocation: the former describes agent competencies and access, while the latter assigns capabilities through ρ(·).Role allocation distinguishes functional positions from the tasks agents execute.
  • Role capability types: Role capabilities are classified as homogeneous or heterogeneous according to whether agents share similar capabilities or have asymmetric capabilities and information access.Homogeneous teams typically use the same underlying LLM, whereas heterogeneous teams may use different models or specialized responsibilities.
  • Homogeneous Role Capabilities: Homogeneous roles reduce engineering complexity and simplify coordination by minimizing capability mismatches, but may be inadequate for highly complex or specialized tasks.Specialized tasks may require verification, domain expertise, or advanced reasoning that homogeneous teams lack.
  • Heterogeneous Role Capabilities: Heterogeneous roles enable division of labor, modular interaction, and specialization, consistently improving task performance, robustness, and adaptability across diverse domains.They can vary in reasoning ability, observation access, permitted actions, or task-specific responsibilities.
  • Role allocation: Role allocation divides labor by assigning tasks and clarifying responsibilities, with static and dynamic approaches determined by how roles are distributed via ρ(·).Static allocation specifies roles and responsibilities before execution, commonly through manually designed prompt templates, and organizes collaboration in structured environments.

3.2 Communication

Communication in LLM-based multi-agent systems combines explicit message exchange with implicit coordination inferred from observations, while protocols govern how agents exchange information and coordinate actions. Explicit communication remains dominant, but adaptive message design and implicit mechanisms extend coordination efficiency, scalability, and robustness when messaging is costly or unavailable.

  • Communication Modes: Communication modes distinguish explicit message exchange from implicit coordination inferred through environmental states, behavioral patterns, perception, memory, and reasoning.Explicit communication uses structured or unstructured messages, whereas implicit communication conveys information through observed actions and environmental interaction.
  • Explicit Communication: Message representations and exchange policies improve communication efficiency by filtering, summarizing, compressing, and selectively transmitting task-relevant information across interaction rounds.AgentPrune reduces redundant interactions, while TodyComm adaptively determines communication targets and message contents.
  • Explicit Communication: Explicit communication dominates LLM-based MASs because it offers transparency, controllability, and compatibility with role-based collaboration.Frameworks combine structured protocols with flexible dialogue mechanisms to improve coordination efficiency and scalability in complex agent ecosystems.
  • Implicit Communication: Implicit communication enables agents to infer intentions, align strategies, and synchronize actions without direct message passing, including through shared environmental cues.Such mechanisms support cooperation when explicit messaging is costly or unavailable, as demonstrated in observation-driven and communication-free coordination settings.
  • Communication Protocols: Communication protocols define rules for information exchange, action coordination, and shared goals, and are categorized as internal protocols or external protocols.Internal protocols govern interactions among agents, whereas external protocols regulate interactions with tools, services, or user interfaces; common standards support heterogeneous agents and large-scale ecosystems.

3.3 Orchestration

Orchestration defines the global coordination topology that organizes agents, information flow, and decision-making in multi-agent systems. Existing work distinguishes centralized, distributed, and hybrid topologies, which trade off control, scalability, autonomy, robustness, and flexibility.

  • Orchestration concept: Orchestration models multi-agent coordination as a topology G = (A, Ec) governing agent connections, information flow, and coordination decisions.It determines how roles and communications integrate into a coherent collaboration process.
  • Orchestration concept: Routing dynamically selects which agent handles a subtask or message, whereas orchestration defines the broader coordination topology in which routing operates.Routing may rely on capability matching, contextual reasoning, or learned policies, and can be locally determined in decentralized architectures.
  • Topological paradigms: Existing orchestration research centers on centralized, distributed, and hybrid topologies that differ in decision authority, coordination responsibilities, information flow, scalability, adaptability, and control.These paradigms organize how agents coordinate and distribute responsibility across the system.
  • Topological paradigms: Centralized orchestration uses a global controller for aggregation, planning, and routing, while distributed orchestration assigns authority locally through peer-to-peer interactions.Centralization simplifies global control but may limit scalability and create single points of failure; distributed approaches support scalable collaboration, autonomy, and robustness.
  • Topological paradigms: Hybrid orchestration combines centralized strategic coordination with decentralized execution, typically using hierarchical structures to balance control with operational flexibility.Higher-level components coordinate goals and planning, while lower-level agents retain execution autonomy.

3.4 Interaction

Interaction concerns the dynamic coordination processes through which agents exchange information, update shared states, and jointly complete tasks. It is organized by information-flow patterns and interaction goals, balancing sequential interpretability and parallel diversity with coordination and aggregation challenges.

  • Interaction captures how agents exchange information, update intermediate states, and jointly drive task completion under a collaboration policy.
  • Information Flow: Sequential interaction passes each agent’s output to the next, supporting modular, traceable workflows but risking error propagation and reduced concurrency.It suits tasks decomposed into dependent subtasks and makes processes easier to debug and optimize.
  • Information Flow: Parallel interaction lets agents explore solution spaces concurrently and aggregate outputs, improving robustness and efficiency while complicating consistency and conflict resolution.Effective mechanisms for reconciling conflicting outputs remain a key research problem.
  • Interaction Patterns: Cooperative interaction aligns agents around shared objectives through structured communication, role division, synchronized decisions, and iterative collective refinement, whereas competitive interaction advances conflicting objectives through negotiation, confrontation, or persuasion.Cooperative systems also use mechanisms such as zero-shot teammate-intention inference, multi-hop communication, feedback loops, and preferential voting.

3.5 Evaluation

Evaluation of multi-agent systems increasingly relies on high-quality, trajectory-level datasets that capture realistic interaction trajectories and diverse collaboration patterns. Recent benchmarks expand from information sharing to dynamic, cooperative, competitive, domain-specific, and long-horizon coordination, enabling evaluation and behavioral refinement.

  • Evaluation foundations: High-quality evaluation datasets should capture realistic interaction trajectories and diverse collaboration patterns.Trajectory-level data is presented as critical for optimizing multi-agent systems.
  • Dynamic evaluation: Recent benchmarks increasingly evaluate dynamic multi-agent systems under incomplete information and sequential decision-making.LLM-Coordination establishes an early collaboration benchmark, while LLMArena adds dynamic environments with incomplete information.
  • Capability coverage: BattleAgentBench measures cooperative and competitive capabilities across zero-sum and non-zero-sum game scenarios.It provides quantifiable metrics for both interaction settings.
  • Orchestration evaluation: MultiAgentBench evaluates orchestration quality through milestone-based KPIs and examines how communication topologies affect collaboration.The benchmark explicitly considers graph-based communication structures.
  • Realistic evaluation settings: Advanced testbeds assess multi-round coordination in high-fidelity medical, enterprise, and real-world planning settings.These settings cover sequential decision-making, information gathering, dispute resolution, end-to-end task completion, coordination efficiency, and long-horizon reasoning.
  • Overall assessment: Trajectory-level benchmarks better reflect MAS interaction and collaboration while enabling accurate evaluation and informed refinement of agent behaviours.This conclusion links benchmark realism to both assessment and behavioral improvement.

3.6 Discussion

The discussion identifies a shift toward dynamic, policy-driven collaboration, while emphasizing that specialization, adaptive coordination, communication efficiency, and trustworthy reasoning remain unresolved design challenges. It frames future multi-agent systems around online adaptation, compact coordination, and fault-aware evaluation.

  • Role design: Heterogeneous agents enable functional specialization and structured cooperation, but increasing role diversity heightens interface mismatches, cascading errors, and coordination overhead.Useful specialization requires reliable, legible outputs and principled role boundaries with capability calibration across agents.
  • Role design: Dynamic role allocation better supports non-stationary and long-horizon tasks by adapting roles to task state, uncertainty, and resource constraints.Current systems often choose dynamically instantiated, reassigned, or retired roles heuristically rather than learning when specialization, redundancy, or additional agents are beneficial.
  • Communication: Explicit communication remains dominant and interpretable, but repeated long textual exchanges increase token overhead, latency, cognitive inefficiency, and inconsistency.Implicit coordination through latent state sharing or memory traces could reduce communication cost, provided interpretability is preserved.
  • Orchestration: No orchestration topology is universally superior: centralized control improves oversight, whereas hybrid and adaptive topologies balance structure with localized autonomy.Future systems should evolve coordination graphs with task complexity, uncertainty, and agent performance rather than treating topology as fixed.
  • Execution: Sequential execution promotes coherence but risks error propagation, while parallel execution improves diversity and robustness but creates costly aggregation conflicts.Future systems will likely combine sequential and parallel execution according to task requirements.
  • Research agenda: Future MAS research should prioritize adaptive coordination, collaboration efficiency, and trustworthy reasoning to balance performance with cost, scalability, reliability, and interpretability.The proposed directions include fewer messages and agents, compact shared representations, interpretable traces, fault localization, and robust evaluation protocols.

4 Multi-Agent System Failure Attribution

Failure attribution in multi-agent systems must trace how localized errors propagate through tightly coupled interactions into system-level failures. The section frames complete execution trajectories as the basis for attribution and argues for methods that explain, verify, and support repair.

  • Failure Attribution Challenges: Localized errors can propagate, accumulate, and evolve across interactions, obscuring their original causes and making system-level anomaly attribution substantially harder.Cascading failures may cause downstream symptoms to be mistaken for root causes, especially in compound or delayed failures.
  • Failure Attribution Definition: Failure attribution uses complete execution trajectories to identify responsible agents, critical steps, failure categories, etiologies, and propagation mechanisms.The goal is to determine who introduced the failure, where it originated, and why it precipitated system failure.
  • Execution Trajectories: A multi-agent execution trajectory records contextual information, state transitions, communications, tool use, decisions, and environmental feedback from initial configuration to terminal state.This trajectory provides the foundation for analyzing collaborative behavior, propagation paths, and failure root causes.
  • Attribution Methods: Reliable attribution cannot rely solely on general-purpose models reading long trajectories; stronger directions reuse failure patterns, model cross-trajectory structure, and support online intervention.Incomplete observability and poor debugging of complex trajectories further undermine attribution quality.
  • Future Directions: Future attribution should integrate attribution semantics, propagation, diagnostic granularity, realistic data, and repair mechanisms while moving beyond localization toward explanation, verification, and repair.Constraint-guided and causal methods increasingly emphasize evidentiary grounding, intervention-based validation, and explaining why anomalies emerge and propagate.

5 Multi-Agent System Self-Evolution

Multi-agent self-evolution extends beyond parameter updates to targeted structural and behavioral adaptation, but current methods face persistent challenges in attribution, retention, scalability, and safety. Progress therefore requires lifelong, hybrid evolution that balances adaptive exploration with verification and alignment.

  • Evolution Foundations: Error attribution enables targeted self-evolution, avoiding computationally inefficient random structural changes and arbitrary prompt adjustments.By pinpointing failure root causes, attribution supplies the context needed for the MAS learning loop.
  • Evolution Targets: Self-evolution jointly adapts agents, communication protocols, orchestration topologies, and behavioral policies beyond parametric updates.This formulation provides a unified language for structural and behavioral adaptation.
  • Evolution Lifecycles: Current evolutionary lifecycles separate ephemeral task-specific structures from reusable cross-task design knowledge, leaving their integration unresolved.Systemic methods construct and dissolve task-specific structures, whereas meta methods build reusable design repositories; bridging these lifecycles remains an open challenge.
  • Lifelong Evolution: Persistent lifelong learning is needed to accumulate collaborative knowledge across task episodes, but catastrophic forgetting, retrieval degradation, and historical-context costs remain major obstacles.Hybridizing parametric and structural evolution, including rapid online topology changes and slow offline parametric consolidation, is proposed to address these limitations.
  • Scalability, Safety, and Alignment: Evolution must balance exploration depth against execution cost while preventing unsafe behavior and guardrail bypass as MAS autonomously rewrite prompts, protocols, and agent populations.EA and MCTS struggle to scale because evaluating candidate architectures requires complete multi-agent workflows, while parametric generators weaken structured-search verification.

6 Conclusion

The survey unifies LLM-based multi-agent systems through the LIFE progression and makes causal dependencies across its stages explicit. It argues that this cross-stage view supports collective intelligence emerging from agent interactions beyond individual capabilities.

  • 6 Conclusion: The survey organizes LLM-based multi-agent systems around LIFE: Lay capabilities, Integrate agents, Find faults, and Evolve through self-improvement.It provides systematic taxonomies for each stage.
  • 6 Conclusion: Collaborative structures determine observable failures, attribution narrows targeted-improvement search, and evolutionary gains reshape collaboration.These dependencies link the four stages into a cross-stage progression.
  • 6 Conclusion: This cross-stage perspective distinguishes the survey from prior work addressing capability, collaboration, attribution, and evolution independently.The survey explicitly contrasts its unified treatment with work that treats these topics separately.
  • 6 Conclusion: The longer-term promise of multi-agent systems lies in cultivating collective intelligence whose system-level capabilities emerge from interactions and exceed any constituent agent’s abilities.The LIFE framework provides a foundation by making these cross-stage dependencies explicit.
Loading 2605.14892v2…