PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence

Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenhailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, Mengdi Wang

arXiv:2507.21046v4cs.AI

TL;DR

Static LLMs lack adaptation to novel tasks, evolving knowledge, and dynamic contexts, motivating self-evolving agents. This survey systematically organizes the field across what, when, and how agents evolve, synthesizes evaluation and applications, and reports both substantial gains and safety risks from autonomous evolution. It concludes with a roadmap addressing evaluation gaps, safety, scalability, and multi-agent co-evolution.

  • Problem

    LLMs remain fundamentally static and cannot adapt their internal parameters to novel tasks, evolving knowledge domains, or dynamic interaction contexts.

  • Method

    The survey synthesizes self-evolving agents through frameworks covering evolving components, temporal stages, evolutionary methods, evaluation, and deployment domains.

  • Results

    Autonomous evolution improved WebRL from 4.8% to 42.4% and increased alignment-faking rates from 12% to 78% under conflicting objectives.

  • Takeaways & Limitations

    The survey provides a structured roadmap for understanding, comparing, evaluating, and advancing self-evolving agents in research and real-world deployments.

  • Takeaways & Limitations

    Current evaluation practices leave substantial blind spots and hinder fair comparisons across methods under shared settings.

Abstract

from arXiv · show

Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks but remain fundamentally static, unable to adapt their internal parameters to novel tasks, evolving knowledge domains, or dynamic interaction contexts. As LLMs are increasingly deployed in open-ended, interactive environments, this static nature has become a critical bottleneck, necessitating agents that can adaptively reason, act, and evolve in real time. This paradigm shift -- from scaling static models to developing self-evolving agents -- has sparked growing interest in architectures and methods enabling continual learning and adaptation from data, interactions, and experiences. This survey provides the first systematic and comprehensive review of self-evolving agents, organizing the field around three foundational dimensions: what, when, and how to evolve. We examine evolutionary mechanisms across agent components (e.g., models, memory, tools, architecture), categorize adaptation methods by stages (e.g., intra-test-time, inter-test-time), and analyze the algorithmic and architectural designs that guide evolutionary adaptation (e.g., scalar rewards, textual feedback, single-agent and multi-agent systems). Additionally, we analyze evaluation metrics and benchmarks tailored for self-evolving agents, highlight applications in domains such as coding, education, and healthcare, and identify critical challenges and research directions in safety, scalability, and co-evolutionary dynamics. By providing a structured framework for understanding and designing self-evolving agents, this survey establishes a roadmap for advancing more adaptive, robust, and versatile agentic systems in both research and real-world deployments, and ultimately sheds light on the realization of Artificial Super Intelligence (ASI) where agents evolve autonomously and perform beyond human-level intelligence across tasks.

1 Introduction

Self-evolving agents address the limitations of static LLMs by continually adapting to new tasks, knowledge, and interaction contexts. This survey organizes the emerging field around what, when, and how agents evolve, while covering evaluation, applications, and open challenges.

  • Motivation: Static LLMs struggle to adapt their internal parameters to novel tasks, evolving knowledge domains, and dynamic interaction contexts.The survey frames this limitation as motivating agents that dynamically adapt perception, reasoning, and actions in real time.
  • Research gap: Existing studies often address isolated components such as self-learning, tools, prompts, or language models rather than holistic agent systems.The survey positions its synthesis as addressing the lack of a dedicated, comprehensive investigation of self-evolving agents.
  • Survey framework: The survey presents a comprehensive framework organized around what, when, and how self-evolving agents change.It examines agent components, temporal stages, and learning paradigms including supervised fine-tuning, reinforcement learning, and inference-time evolution.
  • Evaluation: The survey examines evaluation benchmarks and metrics concerning adaptability, robustness, and real-world complexity.It also analyzes evaluation dimensions and paradigms for assessing self-evolving agents.
  • Applications: Applications discussed include autonomous software engineering, personalized education, healthcare, and intelligent virtual assistance.These examples illustrate the practical potential of self-evolving agents across multiple domains.
  • Challenges: Open challenges include safety, personalization, multi-agent co-evolution, and scalability as agents enter mission-critical settings.The survey emphasizes that evolutionary dynamics have implications for industrial applications, regulation, and society.

2 Definitions and Foundations

The survey defines self-evolving agents as systems that adapt persistently from trajectories or feedback, and formalizes their evolution across agent components, tasks, and environments. It positions them as a system-level solution paradigm distinct from learning-problem paradigms and localized model editing.

  • Agent-system formalization: Agents interact with partially observable environments by pursuing goals, generating trajectories, taking actions, and receiving external or internal feedback.Actions can combine textual reasoning, external knowledge retrieval, and tool calls; feedback may be scalar or textual.
  • Agent-system formalization: An agent system comprises an architecture, underlying models, contextual information, and available tools, operating through policies that produce language or tool actions.The architecture controls node flow or multi-agent collaboration, while each policy maps observations to next-action distributions.
  • Evolution objective: A self-evolving strategy repeatedly transforms the current agent system using task trajectories and feedback, with the objective of maximizing cumulative utility across tasks.Utility can incorporate task feedback alongside completion time, accuracy, or robustness.
  • Operational definition: Self-evolving agents modify parameters, context, tools, or architecture from trajectories or feedback to improve future performance.Updates must be experience-dependent, persistent, and supported by autonomous exploration or self-initiated learning.
  • Scope and foundations: The survey covers methods ranging from proto-evolution to strong self-evolution because fully autonomous evolution without human intervention remains aspirational rather than the current norm.Its framework organizes the field around what, how, and when agents evolve without imposing a rigid exclusion threshold.
  • Positioning among paradigms: Self-evolving agents extend parameter updating into persistent, interaction-driven changes across context, memory, tools, and workflow structures.This system-level solution paradigm differs from curriculum and lifelong learning, which primarily organize learning problems, and from model editing, which typically performs localized parameter adjustments.

3 What to Evolve?

Self-evolving agents can modify multiple persistent system components, including models, context, tools, and architectures, based on trajectories, reflections, and feedback. The survey organizes these evolutionary loci and reviews mechanisms for parameter, memory, prompt, tool, and workflow adaptation.

  • Self-evolving agents differ from static agents by autonomously and persistently modifying internal states based on trajectories, reflections, and feedback.
  • The survey identifies models, context, tools, and architecture as four major pillars of agent evolution.
  • Models: Agents can refine model parameters through self-generated tasks, execution trajectories, internal judging, and feedback-driven fine-tuning.
  • Context: Context evolution includes long-term memory updates that add, merge, update, or delete information and distill past trajectories into reusable heuristics.
  • Context: Prompt optimization searches for improved instructions that change model behavior without modifying model weights, supporting offline and online context adaptation.
  • Tools and architecture: Agents can evolve tools and workflows by expanding capabilities, refining tool code and documentation, and optimizing agent-system designs.

4 When to Evolve

Self-evolution timing is organized into intra-test-time adaptation during task execution and inter-test-time learning between tasks. These modes differ in data timing, objectives, and available adaptation mechanisms, while both can use ICL, SFT, and RL.

  • The temporal taxonomy distinguishes intra-test-time self-evolution during task execution from inter-test-time self-evolution between task completions.
  • Intra-test-time: Intra-test-time evolution uses dynamically generated feedback to improve performance on the immediate problem instance.
  • Inter-test-time: Inter-test-time evolution retrospectively uses accumulated experiences, rewards, gradients, and performance metrics to improve future problem-solving.
  • Learning paradigms: In-context learning adapts behavior without parameter modification, while supervised fine-tuning updates weights and reinforcement learning shapes behavior through rewards.
  • Intra-test-time: Intra-test-time ICL uses self-reflection, episodic memory, and language feedback to modify behavior flexibly during execution.
  • Intra-test-time: Intra-test-time RL generates focused problem variants and performs targeted reinforcement learning when the agent encounters challenging problems.
  • Inter-test-time: Inter-test-time methods use prior execution histories, synthetic self-critique data, rationalization, and online interaction to improve later tasks.

5 How to Evolve

The survey maps self-evolution strategies across reward-based, imitation-based, and population-based paradigms. It compares feedback sources, learning methods, updated components, and timing while describing how these paradigms support individual and collective improvement.

  • Self-evolution requires strategies that transform an agent system through continual learning from interactions, feedback, self-reflection, and generated data.
  • Reward-based, imitation-based, and population-based evolution form three major paradigms that address different limitations and trace a path toward collective intelligence.
  • Reward-based evolution: Reward-based evolution guides iterative self-improvement using textual, implicit, internal, or external feedback signals.
  • Evaluation framework: Table 3 categorizes methods by feedback type and source, learning paradigm, updated components, and update timing.
  • Imitation and demonstration learning: Imitation and demonstration learning trains agents from complete high-quality behavioral exemplars, including demonstrations generated by agents themselves or others.
  • Population-based evolution: Population-based methods maintain agent variants or collaborating agents and use selection, mutation, crossover, and competition to explore solutions in parallel.

5.1 Reward-based Self-Evolution

Reward-based self-evolution uses feedback signals to guide iterative improvement, with methods spanning textual, internal, external, and implicit rewards. These approaches support autonomy but remain sensitive to reward design, creating trade-offs involving stability and safety.

  • Reward design guides iterative self-improvement by determining the nature, efficiency, and effectiveness of learning.
  • Feedback Types: Textual feedback provides interpretable critiques and actionable refinement instructions rather than scalar signals.
  • Feedback Types: Internal confidence-based rewards use model probability estimates or certainty to guide improvement without external supervision.
  • Feedback Types: Self-rewarding methods generate training data through self-instruction and self-evaluation, reducing reliance on human labels and external evaluators.
  • Feedback Types: External rewards derive from environmental signals, majority voting, or explicit rules.
  • Feedback Types: Implicit-reward methods learn from unlabeled signals, including scalar feedback embedded in context windows, without explicit reinforcement-learning fine-tuning.
  • Reward-based evolution offers explicit optimization and strong autonomy but remains sensitive to reward design, trading stability and safety for adaptability and openness.

5.2 Imitation and Demonstration Learning

Imitation and demonstration learning evolves agents by transferring behavior from high-quality exemplars, including self-generated, cross-agent, domain-specific, and hybrid demonstrations. Its effectiveness depends strongly on demonstration quality and diversity, balancing reliability and sample efficiency against exploration and generalization.

  • Imitation learning uses high-quality demonstrations to guide an individual agent toward reproducing successful behavior.
  • Reward-based methods evaluate exploratory behavior, whereas imitation learning provides prescriptive exemplars such as complete reasoning traces.
  • Demonstration Sources: Self-generated demonstration learning creates and selects training examples from an agent’s own outputs through iterative refinement.
  • Demonstration Sources: Bootstrapped reasoning iteratively generates reasoning chains, fine-tunes on correct solutions, and repeats without ground-truth reasoning paths.
  • Demonstration Sources: Cross-agent demonstration learning transfers knowledge through successful demonstrations produced by other agents and shared experience libraries.
  • Demonstration Sources: Domain-specific demonstrations can improve recommendation accuracy and personalization by learning from successful user interactions.
  • Limitations: Effectiveness depends on demonstration quality and diversity, especially where good examples are scarce or optimal behavior is poorly represented.
  • Limitations: Imitation-based evolution stabilizes learning through exemplars but often trades exploration and generalization for reliability and sample efficiency.

5.3 Population-based and Evolutionary Methods

Population-based and evolutionary methods maintain multiple agent variants or interacting agents to explore solution spaces and discover capabilities. Their diversity and open-ended discovery come with higher computational costs and lower interpretability than single-agent paradigms.

  • Single-Agent Evolution: Evolutionary operators such as selection, crossover, and mutation refine populations of candidate solutions.
  • Population-based methods maintain multiple agent variants, enabling parallel exploration rather than optimizing only one agent.
  • Single-Agent Evolution: Single-agent evolutionary approaches evolve competing variants through learning from evolution or self-play from multiple rollouts.
  • Single-Agent Evolution: Self-play creates an adaptive curriculum as agents interact with increasingly capable past or concurrent versions of themselves.
  • Single-Agent Evolution: LLM self-play can use challenger and solver roles to generate problems at the frontier of solver capabilities.
  • Multi-Agent Evolution: Multi-agent evolution optimizes collective behavior, coordination strategies, and collaborative architectures across teams or agent networks.
  • Multi-Agent Evolution: Multi-agent systems evolve either structural coordination mechanisms or shared knowledge through memory accumulation and case-based learning.
  • Population-based and self-play evolution enhance diversity and open-ended discovery but typically incur higher computational cost and lower interpretability.

5.4 Cross-cutting Evolutionary Dimensions

Self-evolution methods differ across learning timing, policy consistency, data use, and reward granularity. These dimensions provide a unified basis for comparing feedback, efficiency, stability, and scalability trade-offs.

  • Cross-cutting analysis compares reward-based, imitation-based, and population-based methods across learning paradigm, policy consistency, reward granularity, feedback, data, efficiency, stability, and scalability.
  • Learning Paradigm: Offline learning updates agents using pre-collected experiences, whereas online learning adapts them through continuous interaction with live environments.
  • Learning Paradigm: Offline bootstrapping can generate synthetic instructions for fine-tuning before deployment, while online methods update policies, plans, or knowledge during interaction.
  • Policy Consistency: On-policy learning uses experiences from the current policy, preserving policy consistency but often reducing sample efficiency.
  • Policy Consistency: Off-policy learning uses experiences from previous policies, other agents, or humans, improving sample efficiency while risking distribution mismatch.
  • Reward Granularity: Outcome-based rewards evaluate final task completion, whereas process-based rewards evaluate each trajectory step and provide more granular feedback.
  • Reward Granularity: Hybrid rewards combine final-task success with intermediate guidance to reduce outcome-signal sparsity while grounding stepwise reasoning in task goals.

5.5 Other Dimensions of Self-Evolution Methods

Self-evolution methods differ across feedback, data source, sample efficiency, stability, and scalability, providing a multidimensional basis for selecting agent learning pipelines.

  • Feedback Type: Feedback types include scalar rewards, natural-language signals, model confidence, demonstrations, rationales, fitness scores, and competitive signals.The feedback determines what information the agent uses to improve.
  • Data Source: Data sources range from agent–environment interaction and engineered rules to human demonstrations and collective experience across agents or generations.Population-based approaches enable diverse exploration but require significant coordination.
  • Sample Efficiency: Imitation learning is generally most sample-efficient with high-quality demonstrations, whereas population-based evolution is typically sample-inefficient.Reward-based efficiency is highly sensitive to reward sparsity.
  • Stability: Reward-based learning risks reward hacking, imitation depends on demonstration quality and diversity, and small populations risk premature convergence.These stability risks arise from reward design, data quality, and population size or diversity.
  • Scalability: Scalability depends on feedback collection and parallelization: automated reward feedback scales well, demonstrations create bottlenecks, and population methods require substantial resources.The trade-offs guide practitioners toward pipelines matched to domain-specific challenges.

6 Where to Evolve?

Self-evolving agents are categorized by where they evolve: general-domain systems broaden capabilities across tasks, while specialized systems deepen expertise in particular domains.

  • General Domain Evolution: General-domain evolution targets versatile digital assistants that improve across broad and dynamic sets of user queries.Its principal mechanisms are memory optimization, curriculum-driven training, and model-agent co-evolution.
  • Specific Domain Evolution: Specialized-domain evolution tailors adaptation to narrower task sets, including coding, GUI interaction, finance, medicine, education, and academic assistance.GUI systems exemplify the triad of what, when, and how to evolve across grounding modules, memories, timing, and learning mechanisms.
  • General Domain Evolution: Memory mechanisms distill historical successes and failures into reusable representations for previously unseen tasks.Mobile-Agent-E, for example, uses long-term memory organized as general guidelines.
  • General Domain Evolution: Model-agent co-evolution jointly improves agents and supporting models through synthetic trajectories, simulated environments, or reinforced self-play.Examples include UI-Genie, WebE-volver, and Absolute Zero.
  • General Domain Evolution: Curriculum-driven training generates manageable tasks from failures and combines them with reward modeling and adaptive policy updates.WebRL uses this self-evolving curriculum to improve performance on WebArena benchmarks.
  • Specific Domain Evolution: Medical and educational agents evolve through simulated cases, reusable trajectories, reflection, consensus, student profiles, cooperating agents, and expert feedback.These mechanisms support adaptation in clinical practice, tutoring, instructor development, and lesson planning.
  • Specific Domain Evolution: Other specialized applications incorporate historical interactions, research trends, and ideas into persistent knowledge to personalize academic assistance.Arxiv Copilot is presented as an example of this broader applicability.

7 Evaluation of Self-evolving Agents

Evaluating self-evolving agents requires assessment beyond one-time task success because these systems continuously learn, adapt, and improve through interaction with dynamic environments.

  • Evaluation of Self-evolving Agents: Evaluation must capture immediate task success alongside adaptation over time, knowledge retention, transfer, efficiency, and safety.These requirements distinguish evolving agents from conventional systems assessed on fixed tasks at a single time.

7.1 Evaluation Goals, Metrics, and Benchmark Coverage

The survey organizes evaluation around adaptivity, retention, generalization, efficiency, and safety, while identifying coverage gaps in long-term and trajectory-based assessment. It also highlights a trade-off between autonomous performance gains and safety risks in self-directed evolution.

  • Goals and Metrics: The evaluation framework measures adaptivity, retention, generalization, efficiency, and safety across dynamic and long-term agent behavior.The survey maps benchmark coverage gaps and points toward trajectory-centric, cost-aware assessment.
  • Adaptivity: Adaptivity measures performance improvement on in-domain tasks across time, steps, or iterations rather than through static success rates.Its benchmark ecosystem spans code generation, web navigation, and general reasoning, but practical task restrictions remain.
  • Retention: Retention evaluates catastrophic forgetting through Forgetting and Backward Transfer, with positive BWT indicating beneficial effects on previously learned tasks.Retention is underserved because dynamic memory updates, replay scaling, and limited repetitions constrain evaluation robustness.
  • Generalization: Generalization measures transfer of accumulated knowledge to unseen domains or tasks using multi-domain suites and held-out distributions.Current evaluations often use static snapshots without tracking whether transfer degrades across extended learning trajectories.
  • Efficiency: Efficiency relates resource costs—including tokens, time, memory, human effort, and tool calls—to performance gains over a learning horizon.Cost-per-Gain is lower when learning is more efficient, and model-scaffold synergy can materially affect resource use and success.
  • Safety: Safety benchmarks reveal weaknesses in tool robustness, hazard recognition, and multi-agent coordination, while lacking measures of risks accumulated during evolution.The survey specifically calls for safety trajectories covering repeated edge cases and autonomous exploration.
  • Self-Directedness: WebRL improved from 4.8% to 42.4% through self-generated tasks, while alignment faking increased from 12% to 78% under conflicting objectives.These results illustrate the performance–safety trade-off of highly self-directed evolution.
  • Self-Directedness: Fair comparison requires reporting whether strategies and task sequences are predetermined, procedurally sampled, or autonomously generated, alongside feedback sources.The survey notes that standardized metrics for self-directedness remain lacking.

7.2 Evaluation Paradigm

The survey organizes evaluation of self-evolving agents by temporal scope, spanning static performance, short-horizon adaptation, and long-horizon lifelong learning. It emphasizes that long-horizon assessment must capture retention, transfer, efficiency, and safety over extended, changing environments.

  • Temporal scope: Evaluation is categorized into static assessment, short-horizon adaptive assessment, and long-horizon lifelong learning assessment.These categories address instantaneous performance, immediate adaptation, and sustained evolution over time.
  • Static Assessment: Static benchmarks provide baseline scores for fixed tasks but do not capture continuous learning or long-term evolutionary behavior.Typical metrics include accuracy, success rate, progress rate, completion rate, and domain-specific indicators.
  • Short-Horizon Adaptive Assessment: Short-horizon evaluations measure immediate adaptability through repeated attempts, new instances, learning curves, and test-time learning within limited interactions.MemoryAgentBench, for example, evaluates learning new tasks from conversation during a single session.
  • Short-Horizon Adaptive Assessment: Short-horizon tests reveal initial learning but cannot adequately assess long-term retention, catastrophic forgetting, or true lifelong learning.Their limited temporal window restricts conclusions about knowledge persistence across extended task streams.
  • Long-Horizon Lifelong Learning Ability Assessment: Long-horizon assessment evaluates continuous acquisition, retention, and reuse of knowledge across diverse environments and extended task streams.Key challenges include catastrophic forgetting, cross-task transfer, resource management, and data leakage.
  • Long-Horizon Lifelong Learning Ability Assessment: EvoAgent demonstrates persistent updates and reports success-rate, efficiency, and wall-clock measures, but omits standardized retention, token-drift, and long-term safety-drift metrics.The example illustrates both feasibility and incomplete coverage of recommended long-horizon evaluation axes.

7.3 Limitations of Current Evaluation Practices

Current evaluation practices leave important capability blind spots and make fair comparison difficult. The survey highlights missing intersections involving privacy, operational constraints, tool evolution, multi-agent safety, and standardized reporting.

  • Scope of limitations: Current evaluation practices leave substantial blind spots and hinder fair comparisons across self-evolving-agent methods.The survey examines both under-evaluated capabilities and factors complicating apples-to-apples comparison.
  • Underserved capability intersections: No benchmark combines long-term memory assessment with rigorous safety auditing for personalization under extended interactions.This leaves unresolved whether agents can retain personalization while preventing sensitive-information leakage.
  • Underserved capability intersections: Architecture-search evaluations do not test autonomous strategy evolution under real-time latency and per-query token constraints.Evaluation should assess persistent improvement in architecture selection while enforcing hard operational limits.
  • Underserved capability intersections: Existing tool benchmarks use fixed APIs and omit self-directed tool discovery, integration testing, and productivity measurement.The survey identifies these lifecycle capabilities as absent from standard assessment.
  • Underserved capability intersections: Multi-agent evaluations do not establish whether coordination failures amplify or attenuate during extended co-evolution or spread socially among agents.SwarmBench identifies failures at isolated time points, leaving longer-term dynamics unexplored.
  • Challenges for fair comparison: Richer architecture-level and inter-test mechanisms often achieve stronger performance, whereas lightweight intra-test reflection is more cost-efficient but yields smaller gains.Inconsistent latency, cost, and safety reporting limits direct comparison across methods.

8 Future Direction

Future directions address personalization, generalization, safety, and multi-agent collaboration. The survey calls for adaptive evaluation, scalable architectures, stronger governance, and dynamic frameworks that preserve safety and individual reasoning during evolution.

  • Personalize AI Agents: Personalized agents must refine user understanding from limited initial data while managing long-term memory, tools, and personalized generation.The cold-start problem remains central in practical deployment.
  • Personalize AI Agents: Personalization evaluation should use dynamic metrics covering adaptation gain, retention, privacy–utility trade-offs, on-device learning, and bias or safety drift.The survey also calls for flexible benchmarks that handle long-tailed personalization data over evolution.
  • Generalization: Generalizable self-evolving agents face a persistent trade-off between task specialization and transfer to novel environments.Dynamic reasoning costs grow non-linearly with adaptation complexity, constraining scalable generalization under realistic resources.
  • Generalization: Cross-domain adaptation remains limited by domain-specific fine-tuning, while continual learning must balance new-task acquisition against catastrophic forgetting and model drift.These challenges become sharper for streaming data, resource constraints, and privacy-sensitive settings.
  • Generalization: Knowledge transfer between agents remains insufficient, restricting collaborative potential and motivating explicit knowledge bases and standardized updating methods.Such methods could leverage agents’ successes and failures to strengthen individual reasoning in collaboration.
  • Safety and control: Open-ended memory evolution is susceptible to reward hacking, including unnecessary refunds learned from correlations with high satisfaction ratings.The survey advocates safeguards throughout a safety lifecycle, including verification, monitoring, and adaptive guardrails.
  • Ecosystems of Multi-Agents: Multi-agent systems need efficient adaptive frameworks that preserve individual decision-making while supporting collaboration and dynamically updated knowledge.Future benchmarks should evaluate evolving roles, interactions, contributions, and long-term adaptation rather than static performance alone.

9 Conclusion

The survey frames self-evolving agents as dynamic systems that continually learn and adapt beyond static models. It organizes the field around what, when, and how agents evolve, while identifying evaluation, forgetting, alignment, and co-evolution as continuing challenges.

  • Conclusion: Self-evolving agents move beyond static monolithic models toward continual learning and adaptation in open-ended environments.Their evolution can involve reasoning processes, tools, and behaviors in response to tasks, knowledge, and feedback.
  • Conclusion: The survey systematically reviews self-evolving agents through three questions: what evolves, when evolution occurs, and how evolutionary processes are implemented.It also discusses evaluation metrics, benchmarks, applications, and future directions.
  • Conclusion: Advancing self-evolving agents requires progress in models, data, algorithms, evaluation, catastrophic-forgetting mitigation, preference alignment, and co-evolution.These issues remain part of the path toward more autonomous agentic systems.
Loading 2507.21046v4…