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

5.1 From Attribution to Evolution / 5.2 Formal Definition of MAS Self-Evolution

• The section begins by linking failure attribution to subsequent multi-agent system self-evolution and then formally defining MAS self-evolution.Source: §5 · p. 3

5.3 MAS Self-Evolution Taxonomy

• It taxonomizes MAS self-evolution into agentic, systemic, and meta self-evolution.Source: §5 · p. 3

5.4 Analyzing Evolutionary Dynamics

• The section analyzes the dynamics governing evolutionary processes in multi-agent systems.Source: §5 · p. 3

5.5 Evaluation

• It includes an evaluation of multi-agent system self-evolution.Source: §5 · p. 3

5.6 Discussion

• The section concludes with a discussion of multi-agent system self-evolution.Source: §5 · p. 3

1 Introduction

• Individual LLM agents support reasoning, planning, tool use, and environmental interaction but struggle with sustained coherence, changing conditions, and multi-step coordination.These limitations become especially consequential in autonomous software engineering, scientific discovery, and embodied robotic control.Source: §1 · p. 4
• Multi-agent collaboration addresses single-agent specialization conflicts by organizing role-specialized agents through structured communication, orchestration, and emerging interoperable protocols.The shift is from building capable individuals toward engineering organizational structures that coordinate agents at scale.Source: §1 · p. 4
• Tighter inter-agent coupling makes collective behavior harder to predict, diagnose, and improve because localized hallucinations, routing errors, or tool mistakes can cascade across interaction rounds.Longer trajectories and deeper dependencies obscure the responsible agent and decisive step, making manual attribution inefficient and unscalable [20] [28] [29] [30].Source: §1 · p. 4
• Failure attribution and self-correction are mutually dependent: diagnosis directs improvement, while acting on diagnoses gives attribution practical value [31, 32, 33].The paper therefore calls for a closed-loop framework in which failure diagnosis directly informs structural self-improvement.Source: §1 · pp. 4–5
• Existing surveys separately cover individual-agent capabilities [36], multi-agent collaboration [38] [39] [40] [41], and agent self-evolution [31] [42], while largely omitting the post-deployment failure loop.No existing survey systematically treats multi-agent failure attribution or the complete operational lifecycle from capabilities through collaboration, diagnosis, and structural improvement [20, 27, 33].Source: §1 · p. 5
• The survey’s LIFE perspective connects individual intelligence, collaboration, failure attribution, and self-evolution as four causally linked stages rather than separate topics.This is presented as the first survey covering the complete operational lifecycle of LLM-based multi-agent systems.Source: §1 · p. 5
• Within LIFE, the survey provides stage-wise taxonomies and comparative analyses spanning individual capabilities, collaboration mechanisms, failure-attribution methods, and self-evolution scopes.The synthesis covers core capabilities, organizational mechanisms, methodological families, and hierarchical scopes across the four stages.Source: §1 · p. 5
• LIFE defines an attribution–evolution loop in which diagnosis narrows targeted self-improvement, while collaborative structures shape which failures can be observed and attributed.The survey also identifies boundary challenges, proposes a cross-stage research agenda, and curates a public repository of literature, taxonomies, and comparative tables.Source: §1 · p. 5

2.1 From LLM to LLM-based Agent: An Architectural Overview

• LLM-based agents address the limitations of stateless language models by perceiving environments, pursuing goals, using tools, and adapting through iterative observation, reasoning, action, and reflection.The survey identifies reasoning, memory, planning, and tool use as the four core individual-agent competencies.Source: §2.1 · p. 6
• The agent architecture forms a closed loop in which memory supplies context, reasoning guides tool selection, planning coordinates multi-step actions, and execution feedback updates future decisions.The resulting action policy is factorized into specialized modules rather than a monolithic input-output mapping, enabling structured and interpretable decision-making.Source: §2.1 · pp. 7–8

2.2 Reasoning

• Reasoning is the cognitive foundation of agent decision-making, and external mechanisms enhance it through input-stage enrichment, reasoning-process improvement, and output-stage regulation.This organization treats foundation-model reasoning as a prior while addressing the gap between static parametric knowledge and dynamic task demands.Source: §2.2 · p. 9

2.2.1 Input-Stage Enhancement

• Input-stage enhancement expands accessible information through knowledge and modality augmentation, including retrieval, structured knowledge exploration, generated knowledge repositories, visual features, and interactive visual reasoning.RAG, Self-RAG, SubgraphRAG, Think-on-Graph, Knowledge Card, RuAG, Multimodal-CoT, Visual Sketchpad, and CogCoM exemplify these directions.Source: §2.2.1 · p. 9

2.2.2 Reasoning-Process Enhancement

• Reasoning-process enhancement combines search-space expansion with path verification, while newer learning and routing methods address the scalability and cost of hand-designed reasoning protocols.DeepSeek-R1 develops autonomous extended reasoning through reinforcement learning, CPO distills tree-search preferences, and adaptive routing can let smaller models match larger ones under comparable compute budgets [63].Source: §2.2.2 · p. 10
• Process-level verification improves reasoning reliability by scoring individual steps, addressing the limitation of methods that evaluate only complete outputs.Lightman et al. use human step labels, Math-Shepherd estimates correctness through Monte Carlo rollouts, and WizardMath applies process rewards during both training and inference.Source: §2.2.2 · p. 10

2.2.3 Output-Stage Regulation

• Output-stage regulation combines reliability assessment with response-behavior control to detect factual errors and adapt generation through evidence checking, internal signals, decoding interventions, alignment, refusal, abstention, or tool-based verification.Representative methods include FActScore, SelfCheckGPT, Semantic Entropy, CoVe, RARR, DoLa [79], FactTune, R-Tuning, CRITIC [87], and SEAL [88].Source: §2.2.3 · pp. 10–11

2.3 Memory

• Memory forms a closed loop by recording, consolidating, and recalling task-execution experiences to inform better actions in subsequent interactions.The scope is agent-level experiential memory, distinct from model-level parametric knowledge and static external knowledge bases.Source: §2.3 · p. 11

2.3.1 Memory Formation

• Memory formation ranges from direct recording to abstractive distillation and typed routing, while the formed memory type determines support for self-correction, planning, and cross-task skill reuse.SeCom finds that topically coherent, compression-denoised segments outperform turn-level and session-level alternatives; ENGRAM [116] routes entries into episodic, semantic, or procedural stores, whose collapse substantially degrades retrieval quality.Source: §2.3.1 · p. 13

2.3.2 Memory Maintenance

• Memory maintenance addresses unbounded growth and limited attention through storage structures and dynamic policies that organize, update, compress, forget, or archive memories.Flat append-only storage limits scalability, whereas graph and hierarchical structures add relational or multi-granularity organization; maintenance spans decay-based forgetting, LLM-selected operations, and learned utility policies.Source: §2.3.2 · p. 14

2.3.3 Memory Retrieval and Utilization

• Memory retrieval converts accumulated state into actionable context; dense vector retrieval ranks memory entries by cosine similarity to the current query but ignores temporal and importance signals, whereas multi-factor scoring combines dimensions.Generative Agents weight recency, LLM-assigned importance, and embedding relevance into a composite score.Source: §2.3.3 · pp. 14–15

2.4 Planning

• Planning transforms high-level goals into structured executable actions while handling sub-task dependencies, environmental feedback, and potential failures over extended horizons.Source: §2.4 · p. 15

2.4.1 Decomposition-Based Planning

• Decomposition-based planning organizes complex goals into sub-tasks either proactively before execution or progressively during execution, with explicit proactive planning outperforming reactive approaches on complex web benchmarks [138].Proactive methods include sequential planning, dependency graphs, parallel execution, symbolic planning, and plan skeletons; progressive methods refine plans from observations, failures, or evolving goals.Source: §2.4.1 · pp. 15–16
• Progressive decomposition interleaves planning with execution, refining plans from intermediate observations; ReAct alternates reasoning traces and actions, Reflexion uses critiques of failed episodes, and ADaPT recursively decomposes failed sub-tasks.Source: §2.4.1 · p. 16

2.4.2 Search-Based Planning

• Search-based planning explores multiple candidate paths through evaluation, pruning, or backtracking, using step-level methods such as MCTS and plan-level generation, comparison, retrieval, and experience-augmented selection [96, 127, 144].Applications span household tasks, prompt optimization, world-model reasoning, process-reward-guided search, workflow search, API-chain A*, diffusion-based implicit search, and knowledge-graph reranking.Source: §2.4.2 · p. 17

2.5 Tool Use

• Tool use expands an agent’s action space through external functions, APIs, code interpreters, web services, and systems, requiring structured calls and integration of returned results.The section frames tool use as crossing the boundary between language and execution.Source: §2.5 · pp. 17–18

2.5.1 Tool Capability Acquisition

• Tool capability acquisition requires recognizing when tools are needed, selecting them, generating valid parameters, and incorporating results, using trajectory-based learning or protocol-based alignment.Trajectory-based methods learn from demonstrations, whereas protocol-based methods ground behavior in tool interface descriptions.Source: §2.5.1 · p. 18

2.5.2 Tool Invocation

• Tool invocation progresses from one-shot calls to closed-loop feedback and workflow orchestration, with increasingly explicit control over repeated calls, dependencies, data flow, and parallel execution.Examples include ReAct-style reasoning–action interleaving, LLMCompiler’s dependency-tracked DAGs, and workflow search using MCTS.Source: §2.5.2 · pp. 19–20

2.5.3 Tool Generalization

• Tool generalization addresses unseen-tool adaptation through documentation comprehension and tool discovery through retrieval from large, evolving API collections.GenTool covers zero-to-one and weak-to-strong transfer, while retrieval methods target selecting relevant tools from large pools.Source: §2.5.3 · pp. 20–21
• ToolLLM demonstrates strong zero-shot generalization on the out-of-distribution APIBench benchmark despite training on a disjoint API set.This result illustrates transfer beyond the APIs observed during training.Source: §2.5.3 · p. 20

2.6 Evaluation

• Benchmarks increasingly evaluate reasoning, planning, and tool use jointly, while incorporating trajectories, subgoal completion, and execution traces beyond outcome-based metrics.These shifts provide finer diagnostic resolution for identifying where agents succeed and fail.Source: §2.6 · p. 21
• Comprehensive suites such as AgentBench, GAIA, MINT, AgentBoard, and TheAgentCompany [179] stress-test agents across environments, tools, partial progress, and long-horizon workflows.TheAgentCompany [179] specifically evaluates realistic enterprise workflows requiring long-horizon coordination across web and code.Source: §2.6 · p. 21
• Specialized benchmarks span web, operating systems, software engineering, tool use, planning, and memory, including WebArena, OSWorld, SWE-bench, ToolBench [185], ALFWorld, and LoCoMo.These benchmarks target reproducible web tasks, state verification, real GitHub issue resolution, API use, long-horizon action sequences, and long-conversation memory.Source: §2.6 · p. 21
• Domain-specific and safety-critical evaluations cover scientific research, enterprise applications, cybersecurity, and harmful tool-use behaviors through MLE-bench [193], PaperBench [196], WorkArena++ [197], Cybench [199], and AgentHarm [200].These settings broaden evaluation beyond general-purpose tasks, although WorkArena++ [197] and CRMArena [198] remain only partially open.Source: §2.6 · p. 22
• Current benchmarks rarely measure how improving one capability affects another or incorporate token consumption, latency, and cost, motivating integrated, resource-aware assessment.Single-agent capability bottlenecks also motivate the multi-agent collaboration strategies discussed next.Source: §2.6 · p. 22

2.7 Discussion

• Modular optimization overlooks interdependent reasoning, memory, planning, and tool use, allowing capability mismatches to become deployment bottlenecks invisible to benchmarks targeting only one or two capabilities.The formalization in Section 2.1 makes these dependencies explicit, but existing methods predominantly optimize modules in isolation.Source: §2.7 · p. 22
• Reasoning enhancement, hallucination suppression, and memory design remain fragmented: broader search increases error exposure, process reward models scale poorly, reliability is mostly post-hoc, and memory components lack systematic interaction studies.Graph storage improves relational retrieval but raises maintenance cost [118], utility-driven maintenance needs extensive interaction [126], and storage formats often ignore retrieval compatibility.Source: §2.7 · p. 22
• Agent capabilities are evaluated as static snapshots even though deployment requires continuous adaptation to changing strategies, tools, and memories.Tool generalization and memory evolution are commonly treated as one-time capacities rather than ongoing processes.Source: §2.7 · p. 22

• MASs combine agents with distinct capabilities to tackle tasks beyond the ability of any individual agent [202].Source: §3 · p. 22
• The collaboration framework decomposes a MAS into Role, Communication, Orchestration, and Interaction, each addressing a fundamental question about collaborative intelligence.Source: §3 · p. 24
• Role assigns specialized responsibilities such as planning, reasoning, or execution, while Communication governs how agents generate, structure, and transmit messages from local histories and shared context.Source: §3 · p. 25
• Orchestration specifies agent connectivity and message-passing pathways, while Interaction describes policy-driven execution in which agents act, communicate, and coordinate through trajectories of actions, messages, and reasoning steps.Source: §3 · p. 25

3.1.1 Role Capability

• Role capability captures agents’ competencies, action and observation spaces, information access, and influence over environments or other agents, and is classified as homogeneous or heterogeneous.Homogeneous configurations simplify engineering and coordination but may lack specialized expertise; heterogeneous configurations support division of labor, specialization, robustness, and adaptability across complex tasks.Source: §3.1.1 · pp. 25–27

3.1.2 Role Allocation

• Role allocation divides labor by mapping agent capabilities to task requirements, distinguishing functional roles from the tasks those roles enable agents to perform.This distinction is represented by the mapping ρ : A → Z and organizes responsibility assignment across collaborative systems.Source: §3.1.2 · p. 27
• Static role allocation predefines roles and responsibilities, reducing ambiguity and coordination complexity while improving execution efficiency in structured environments.Its rigidity limits adaptation when task requirements evolve or unexpected situations arise, motivating dynamic alternatives.Source: §3.1.2 · pp. 27–28
• Dynamic role allocation lets agents select, adjust, instantiate, or evolve roles during execution according to task requirements and environmental context.It supports flexible, scalable collaboration in open-ended settings but requires continual state and resource evaluation, adding coordination overhead and potentially reducing stability.Source: §3.1.2 · p. 28

3.2 Communication

• Communication is a fundamental mechanism for coordinating decisions, sharing intermediate results, and advancing shared objectives in multi-agent systems [39].Source: §3.2.1 · p. 29

3.2.1 Communication Modes

• Communication modes comprise explicit exchanges of instructions, results, reasoning, or feedback and implicit coordination through observed behaviors, environmental states, and inferred intentions.Explicit communication remains dominant because of its transparency, controllability, and compatibility with role-based collaboration, while implicit communication can coordinate agents when messaging is costly or unavailable.Source: §3.2.1 · pp. 29–30

3.2.2 Communication Protocols

• Communication protocols define concrete rules for information exchange and divide into internal protocols for agent coordination and external protocols for tools, services, databases, APIs, or interfaces.A2A supports peer-to-peer task handoff, ANP supports service discovery and message routing, and MCP standardizes external resource integration while improving modularity, maintainability, and reusability.Source: §3.2.2 · pp. 30–31

3.3 Orchestration

• Orchestration models multi-agent coordination as a topology G = (A, E_c) that structures agent connections, information flow, and coordination decisions, while routing selects the next agent or subtask.Routing may use capability matching, contextual reasoning, or learned policies, including locally determined decisions in decentralized architectures [223, 281, 282, 283].Source: §3.3 · pp. 31–32

3.3.1 Centralized Orchestration Topology

• Centralized orchestration gives one controller global decision authority to aggregate information, determine interaction flow, and redistribute instructions, but can limit scalability and create single points of failure.Examples enhance centralized memory, planning, topology reconstruction, routing efficiency, reinforcement-learning control, or constraint management [284, 285, 286, 287, 288, 289].Source: §3.3.1 · p. 32

3.3.2 Distributed Orchestration Topology

• Distributed orchestration removes the global coordinator, enabling peer-to-peer information flow and local routing decisions that support scalable collaboration while preserving agent autonomy and system robustness.Representative systems use dynamic networks, decentralized executor and router pools, learned coordination policies, localized information synchronization, or publish–subscribe communication [283, 290, 291, 292, 293, 294].Source: §3.3.2 · pp. 32–33

3.3.3 Hybrid Orchestration Topology

• Hybrid orchestration combines centralized strategic coordination with decentralized execution, typically using hierarchical structures to balance global goal alignment with operational flexibility.Frameworks assign higher-level planning or aggregation to central or parent nodes while preserving lower-level agent autonomy, and some dynamically select execution topologies [13, 247, 295, 296, 297, 298, 299].Source: §3.3.3 · p. 33

3.4 Interaction

• Interaction governs how agents exchange information, update intermediate states, and jointly drive task completion during execution under πcoll.Source: §3.4 · p. 33

3.4.1 Information Flow

• Sequential interaction offers interpretable, modular, and traceable workflows for dependent subtasks but risks error propagation and reduced efficiency when tasks could run concurrently.Parallel interaction instead enables concurrent exploration that can improve robustness and efficiency, while making aggregation and consistency maintenance challenging.Source: §3.4.1 · p. 34

3.4.2 Interaction Patterns

• Cooperative interaction uses shared objectives, structured communication, role organization, and iterative alignment to improve collective performance and joint-decision robustness.Examples include role-based workflows, inferred teammate intentions, multi-hop communication, shared reflection, and preferential voting.Source: §3.4.2 · p. 35
• Competitive interaction arises from conflicting objectives and drives strategic adaptation through debate, negotiation, persuasion, and adversarial exchanges.Frameworks study bargaining, strategic dilemmas, real-time bidding, role diversity, and game-theoretic preference alignment.Source: §3.4.2 · pp. 35–36

• High-quality evaluation datasets should capture realistic interaction trajectories τ and diverse collaboration patterns, making dataset construction critical for optimizing MAS.Source: §3.5 · p. 36
• Recent benchmarks progress from information sharing and collaboration to dynamic decision-making, cooperative and competitive capabilities, and communication-topology evaluation, including LLM-Coordination, LLMArena, BattleAgentBench [329], and MultiAgentBench.Source: §3.5 · p. 36
• High-fidelity evaluations test sequential decision-making, information gathering, dispute resolution, end-to-end completion, and long-horizon planning in medical, enterprise, and real-world settings [331] [333] [334].Source: §3.5 · pp. 36–37
• Trajectory-level benchmarks better reflect MAS interactivity and collaboration while enabling accurate evaluation and informed refinement of agent behaviors.Source: §3.5 · p. 37

Role configuration

• Heterogeneous specialization enables structured division of labour, but increases interface mismatches, cascading errors, coordination overhead, and the need for principled role boundaries and capability calibration.Specialized outputs must remain reliable and legible to downstream agents.Source: §3.6 · p. 37

Role allocation

• Dynamic role allocation better supports non-stationary and long-horizon tasks by adapting roles to task state, uncertainty, and resources, but current systems rarely learn when specialization or redundancy is beneficial.Static assignment remains effective when task decomposition is predictable and stable.Source: §3.6 · pp. 37–38

Communication

• Explicit communication remains transparent and debuggable, yet scaling agents and interaction rounds makes it costly, verbose, latency-intensive, and vulnerable to inconsistent context interpretation.Future systems should investigate implicit coordination through shared memory, environment state, latent state, or memory traces while preserving interpretability.Source: §3.6 · p. 38

Orchestration topology

• No orchestration topology is universally superior: centralized designs provide oversight but risk bottlenecks, distributed designs improve autonomy but complicate consensus, and hybrid designs combine structure with localized autonomy.The discussion treats topology as an adaptive variable that should evolve with task complexity, uncertainty, and agent performance.Source: §3.6 · p. 38

Execution and interaction

• Sequential execution supports refinement and verification but propagates early errors, whereas parallel execution improves diversity and robustness while creating costly aggregation conflicts; cooperation dominates despite underexplored benefits and risks of competition.Competitive interactions can improve reasoning through adversarial review, debate, and bargaining, but may degenerate without incentives and adjudication.Source: §3.6 · pp. 38–39

Cross-cutting challenges and agenda

• Future MAS research should prioritize adaptive coordination, collaboration-efficient designs, and trustworthy reasoning because larger systems do not guarantee better performance and amplify interaction failures, cost, and reliability challenges.These directions involve online selection of roles, topology, and communication, fewer messages and agents, compact shared representations, interpretable traces, fault localization, and robust evaluation.Source: §3.6 · p. 39

4 Multi-Agent System Failure Attribution

• Propagating errors from decomposition, communication, and coordinated decisions can surface later as system-level failures, obscuring their origins and complicating attribution [20] [342].Tightly coupled interactions produce cascading failures, making manual localization of responsible agents and time steps inefficient and unscalable [20, 343, 344, 345].Source: §4 · p. 39, §4.1 · p. 39

4.1 From Collaboration to Diagnosis

• Failure attribution analyzes a complete execution trajectory to identify responsible agents, critical steps, failure categories, etiologies, and propagation mechanisms [342] [346] [348].The trajectory includes state transitions, communications, tool use, decisions, and environmental feedback, supporting root-cause analysis [347] [348] [349].Source: §4.1 · p. 40

4.1.1 Formal Expression of the Attribution Process

• The formal attribution process returns a responsible agent and decisive time step only when trajectory evaluation marks the final state anomalous; otherwise, it returns ∅.The formulation can extend from single agent-step attribution to multi-agent, multi-step, and causal-chain settings [344].Source: §4.1.1 · pp. 40–41

4.2 Failure Taxonomy

• Failure taxonomies classify anomalies by system structure, execution stage, or causal lifecycle to support attribution, diagnosis, and repair [342] [346].Structural views capture coordination and verification breakdowns, execution views locate reasoning, tool, planning, or module errors, and lifecycle views trace root causes and propagation.Source: §4.2 · p. 41, §4.2.1 · p. 41, §4.2.2 · p. 41

4.2.3 Causal Lifecycle

• Existing taxonomies lack unified generalizability and granularity, while single-cause–single-effect attribution misses coupled, cascading multi-cause–multi-effect anomalies [344, 347, 355].The section calls for composite labels and cross-layer associations to characterize how local triggers evolve into global system dynamics.Source: §4.2.3 · p. 42

4.3.1 Data-driven Method

• Data-driven attribution shifts from manual or generic long-context analysis toward learnable trajectory representations, specialized attribution models, and lightweight diagnostic modules [28] [345] [347].Representative systems construct supervision through counterfactual replay, programmatic or context-aware error injection, and structured error signals, addressing scarce and fragmented failure data.Source: §4.3.1 · pp. 42–43
• Data-driven methods have advanced from direct localization to structured modeling that reuses failure patterns, captures cross-trajectory dependencies, and supports online intervention.Recent systems monitor propagation, issue preemptive alerts, and assist human diagnosis rather than relying solely on general-purpose models reading long trajectories.Source: §4.3.1 · p. 44

4.3.2 Constraint-Guided Method

• Constraint-guided methods decompose attribution into executable, verifiable subtasks that delimit scope, structure error analysis, and validate root-cause judgments with explicit evidence.Their main advantage is transparent, auditable diagnosis, but reliance on task-specific priors, manual constraints, and domain-dependent criteria limits generalizability.Source: §4.3.2 · pp. 44–45

4.3.3 Causal-inference Method

• Causal-inference methods explain why failures emerge and propagate by combining counterfactual effects, execution-log causal graphs, and structured verification of attribution chains.They distinguish genuine root causes from propagated symptoms and model influences transmitted through other agents and environmental transitions.Source: §4.3.3 · p. 46

4.4 Failure Attribution Evaluation

• Failure-attribution datasets increasingly emphasize realistic data, finer-grained annotations, observable processes, and composite evaluation beyond single-metric localization accuracy.Construction has moved toward automated large-scale generation, while labels now include error types, interaction edges, functional modules, and root causes; Table 9 compares these dataset dimensions.Source: §4.4 · p. 47, Table 9 · p. 47
• Failure-attribution datasets increasingly connect workflow configurations, propagation structure, behavioral diagnoses, and evaluation protocols rather than merely expanding sample scale.Research emphasizes observational fidelity, traceable synthetic labels, real-failure alignment, and richer attribution targets, while balancing scalability against faithful failure representation.Source: §4.4 · p. 48

4.5.1 Core Challenges

• Multi-agent anomaly attribution remains difficult because its semantics are unsettled, with studies variously targeting agents, steps, error types, interactions, workflows, or multi-perspective responsibility objects.This inconsistency hinders direct method comparison, annotation alignment, transfer, data reuse, and benchmark unification, while conflating manifestations, triggers, and responsibility bearers weakens credibility.Source: §4.5.1 · p. 49
• Failures propagate and amplify across agents, modules, and stages, so observable symptoms often differ from origins and current approaches can miss compound, delayed, and multi-error causes.Fine-grained attribution further suffers from long-horizon dependencies, context sensitivity, redundancy, noise, and weakened long-range signals, limiting reliable downstream repair.Source: §4.5.1 · pp. 49–50
• Attribution data remain insufficiently realistic because automated injection and counterfactual samples differ from deployment failures, whereas high-fidelity manual annotation is costly, hard to scale, and disagreement-prone.The central requirement is simultaneous realism, scalability, and consistency, supported by stronger observational fidelity and unified attribution granularity and evaluation standards.Source: §4.4 · p. 48, §4.5.1 · p. 50
• Evaluation still centers on localization accuracy and is weakly linked to verification, intervention, or repair, leaving attribution without a robust closed loop from diagnosis to system improvement.Because tasks increasingly involve multi-cause, cross-layer, and multi-granularity failures, evaluation must assess reasoning consistency, ranking quality, plausibility, and repair usefulness.Source: §4.4 · p. 48, §4.5.1 · p. 50

4.5.2 Future Research Direction

• Future attribution systems should unify classification, model propagation and causal responsibility, improve realistic granular datasets, and establish multidimensional attribution–verification–repair pipelines.The intended progression is from localizing faults toward explaining, verifying, and translating diagnostic evidence into practical repair gains.Source: §4.5.2 · p. 51

5.1 From Attribution to Evolution

• Self-evolving MAS use recognized failures to search for improved collaborative states, addressing architectural rigidity and scalability bottlenecks that limit fixed orchestration and rule-based adaptation.Attribution narrows the evolutionary search space, while growing topology and policy spaces make manual orchestration increasingly ineffective.Source: §5.1 · p. 52

5.2 Formal Definition of MAS Self-Evolution

• Self-evolution is formalized as a generation-indexed transition that uses historical trajectories, error attributions, and peer critiques to jointly adapt agents, protocols, topologies, and policies for expected utility across task distributions.The framework distinguishes intrinsic improvement from environmental variation by evaluating objectives over diverse task distributions.Source: §5.2 · pp. 53–54

5.3 MAS Self-Evolution Taxonomy

• The taxonomy separates self-evolution into agentic changes to prompts, memory, or parameters; systemic restructuring of collaboration; and meta-level selection among complete system designs.These scopes progress from individual capability refinement to organizational restructuring and transferable design knowledge across diverse tasks.Source: §5.3 · p. 54

5.3.1 Agentic Self-Evolution

• Prompt evolution refines textual instructions, memory evolution accumulates reusable experience, and parameter evolution internalizes capabilities through RL or SFT while avoiding increased runtime inference costs.Examples include shared reflection, adversarial self-play, interaction rewards, and co-evolutionary safety training; parameter evolution is computationally expensive to train but embeds capabilities in model weights.Source: §5.3.1 · p. 55
• Prompt evolution changes textual prompts while underlying model parameters remain frozen; memory evolution builds persistent experience banks for online-persistent learning but shifts bottlenecks toward retrieval efficiency and homogenized historical biases.Source: §5.3.1 · p. 55
• Agentic self-evolution improves individual agents through peer interactions, including adversarial self-play [394], mutual critique [387], and collaborative knowledge sharing [397].These interactions provide autonomous feedback signals that support internal adaptation within multi-agent contexts.Source: §5.3.1 · pp. 55–56

5.3.2 Systemic Self-Evolution

• Systemic self-evolution targets communication topology, team composition, and shared memory, distinguishing ephemeral task-specific structures from persistent organizational knowledge.Topology and composition adaptations are typically constructed and dissolved per task, whereas shared memory accumulates knowledge across execution episodes; bridging these lifecycles remains open.Source: §5.3.2 · pp. 56–57

5.3.3 Meta Self-Evolution

• Meta self-evolution treats complete system specifications as candidates, building reusable design knowledge across tasks through either archive-based search or faster learned generators.Archive-based approaches offer broader exploratory coverage, while generators compress design knowledge for speed; combining both strategies remains an open question.Source: §5.3.3 · p. 58

5.4 Analyzing Evolutionary Dynamics

• Across evolutionary loci, six mechanisms are analyzed with the Variation–Selection–Retention framework, while LLM reflection risks self-referential selection and archival saturation.Objective compiler feedback [403] can decouple generation from evaluation, but reflection remains accessible because it requires no training infrastructure and is bounded by the underlying model’s generative capacity.Source: §5.4 · p. 59
• Evolutionary mechanisms trade off diversity, targeted diagnosis, retention, and scalability: EA preserves diverse candidates, textual gradients enable causal structural edits, while RL and search face costly evaluations.EA-based methods support multi-objective capability–safety selection and behavioral diversity, whereas textual gradients identify failed subtasks and missing capabilities but retain knowledge only ephemerally. RL can encode gains in model weights without inference overhead but risks catastrophic forgetting and reward misspecification.Source: §5.4 · pp. 60–61
• Table 13 shows that RL, LLM Reflection, and EA span all three evolutionary loci, whereas textual gradients and heuristic updates remain systemic-only, with hybrid mechanisms rare.The surveyed landscape contains 31 frameworks and 36 mechanism–locus circles; the gaps suggest opportunities for broader hybridization.Source: §5.4 · p. 61, Table 13 · p. 62

5.5 Evaluation

• Evaluation spans static benchmarks, interactive environments, embodied sandboxes, and adversarial arenas, with static tests measuring capability gains but missing dynamic adaptability.Interactive testbeds assess multi-step workflow success and collaboration stability, embodied environments probe long-term survival and shared memory, and adversarial arenas detect unsafe or deceptive behaviors.Source: §5.5 · p. 62

5.6 Discussion

• Existing online-ephemeral and offline-persistent lifecycles fall short of continuous lifelong adaptation, motivating MAS that accumulate collaborative knowledge without resetting organizational memory.Parametric evolution risks catastrophic forgetting, while explicit memory banks suffer declining retrieval accuracy and increasing historical-context costs; hybrid parametric–structural evolution and dual-process architectures are proposed.Source: §5.6 · p. 63
• Combinatorial architecture spaces and full-workflow fitness evaluations make self-evolution computationally expensive, causing EA and MCTS to struggle with meta-level structural discovery [396] [407].The central challenge is balancing evolutionary exploration depth against inference time and API-token costs.Source: §5.6 · p. 63
• Efficient architectural generators offer a route to scalable evolution by learning distributions of effective topologies and sampling task-specific architectures in one step.MAS-GPT [411] and Agentic Supernets [410] exemplify this generative modeling direction.Source: §5.6 · p. 63
• Hierarchical optimization should combine lightweight generator-driven real-time adjustments with rigorous offline structured searches such as MCTS.Parametric generators enable fast adaptation but sacrifice the verification guarantees of structured search.Source: §5.6 · p. 63
• Self-evolving MAS can bypass static safeguards, reward-hack objectives through deception or collusion, and therefore require evolutionary alignment as a continuous selection pressure [407].Traditional RLHF is designed for isolated foundation models and is vulnerable when transferred to self-evolving multi-agent contexts.Source: §5.6 · pp. 63–64
• Static benchmarks such as MMLU [415] and GSM8K [416] cannot assess structural adaptation, shifting-task resilience, or resource-constraint robustness in evolving MAS.Standardized open-ended protocols should measure evolutionary sample efficiency, behavioral diversity, and archives of specialized topologies in embodied and software-engineering environments [397, 403, 423].Source: §5.6 · p. 64

6 Conclusion

• The LIFE progression organizes systematic taxonomies and comparative analyses of individual capabilities, collaboration mechanisms, failure attribution, and self-evolution.It covers four stages: Lay, Integrate, Find, and Evolve.Source: §6 · p. 64
• Collaborative structures determine observable failures, attribution narrows targeted improvement, and evolutionary gains reshape the collaborative fabric.These causal dependencies distinguish the survey from prior work addressing the topics independently.Source: §6 · p. 64
• The longer-term promise of multi-agent systems is collective intelligence whose system-level capabilities emerge from interactions and exceed any constituent agent’s abilities.The LIFE framework provides a foundation for making these cross-stage dependencies explicit.Source: §6 · p. 64