Self-Evolution Method

• Self-evolution method is an iterative paradigm where systems acquire, refine, update, and evaluate their own experience to pursue autonomous improvement.
• It encompasses dynamic modifications to various artifacts such as memory, code, and context across domains like embodied AI and software agents.
• The approach bridges adaptation and evolution by transforming a system's operational design domain to meet new requirements beyond its original configuration.

Self-evolution method denotes an iterative autonomous-improvement paradigm in which a model, agent, or larger system acquires, refines, and reuses experience generated during its own operation to improve future behavior. In the survey formulation, the evolving process is an iterative cycle composed of experience acquisition, experience refinement, updating, and evaluation (Tao et al., 2024). Across recent work, the evolved object is not fixed: depending on the domain, self-evolution may target memory, task objectives, world models, embodiment assumptions, prompts and contexts, executable code, diagnostic tools, team organization, or complete interaction trajectories (Feng et al., 4 Feb 2026, Robol et al., 29 Apr 2026, Hao et al., 28 May 2026). This suggests that the term refers less to a single trainable algorithm than to a family of closed-loop designs for continual self-modification under feedback.

1. Conceptual boundary between adaptation and evolution

A central distinction in the literature is the difference between self-adaptation and self-evolution. In the operational-design-domain formulation, self-adaptation means changing configurations or behavior so that the system continues to operate within its existing ODD, whereas self-evolution means altering the system so that its ODD itself changes—typically by extension—allowing it to satisfy requirements in contexts not covered by the initial design (Weyns et al., 2023). The ODD is defined as the conditions for which the system was built, including requirements $R$, constraints $B$, and context $C$, and the paper explicitly represents evolution as an ODD transformation $ODD_{Se} = ODD_{Si} \cup ODD_e$ (Weyns et al., 2023).

The same distinction appears in software-agent work. Traditional autonomous agents may update beliefs, choose among existing desires, and execute available plans, but remain “bound to requirements, goals, and capabilities fixed at design time.” Self-evolving software agents instead can autonomously revise goals, reasoning structures, executable actions, plans, and code, with an Automated Evolution Module triggered when perceived information cannot be interpreted or exploited using current knowledge and goal structures (Robol et al., 29 Apr 2026).

Embodied-AI work sharpens the contrast further. Conventional embodied AI is described as operating in a human-crafted setting with given memory, given tasks, fixed embodiments, and a specified environment, optimized by a pretrained model. Self-evolving embodied AI shifts from that fixed setting to a system with memory self-updating, task self-switching, environment self-prediction, embodiment self-adaptation, and model self-evolution, all coupled over time (Feng et al., 4 Feb 2026). A recurring misconception is therefore that self-evolution is merely stronger adaptation. The literature instead treats it as change in the system’s own design space, objective structure, or internal representational apparatus.

2. Shared architectural pattern

Although implementations differ, many self-evolution methods instantiate a recurring closed loop. The survey view is abstract—acquire experience, refine it, update, evaluate—whereas domain-specific papers specialize the loop around their own evolving artifacts (Tao et al., 2024). In embodied AI, the agent maintains a self-state consisting of own physical state, internal cognition state, and external world state; task self-switching determines objectives, memory self-updating curates relevant experience, model self-evolution updates the learning machinery, the agent acts, and feedback updates cognition, embodiment, and world models before the loop repeats over time (Feng et al., 4 Feb 2026). In multi-agent systems, Meta-Team runs an episode-by-episode loop: execute a task, freeze the execution experience $e_k=(x_k,\tau_k,r_k)$, run post-task collaborative evolution, commit validated scaffold updates, and reuse the evolved MAS on future tasks (Hao et al., 28 May 2026). In software agents, the BDI loop is kept distinct from an Automated Evolution Module that is invoked only when current knowledge, goals, or actions are semantically insufficient (Robol et al., 29 Apr 2026).

The recurrent pattern can be summarized by the kinds of artifacts that are allowed to change.

Setting	Evolving artifact	Representative mechanism
Prompt-based LLMs	Instruction field or context	Context edits plus tree-guided evolution loop
Embodied agents	Memory, tasks, world model, embodiment, model	Five coupled self-evolution modules
Multi-agent systems	Agent scaffold, pairwise coordination, team constitution	L1/L2/L3 collaborative post-task evolution
Software agents	Goals, reasoning, executable code	BDI loop plus Automated Evolution Module
Open-world embodied control	Skills and guardrails in external knowledge	Diagnosis plus dual-track distillation

These representative instances are explicitly described in work on Learning to Self-Evolve, self-evolving embodied AI, Meta-Team, self-evolving software agents, and Steve-Evolving (Chen et al., 19 Mar 2026, Feng et al., 4 Feb 2026, Hao et al., 28 May 2026, Robol et al., 29 Apr 2026, Xie et al., 13 Mar 2026). What varies is not the existence of a loop, but the locus of state change and the mechanism by which experience is turned into persistent modification.

3. Major methodological families

One major family evolves context, prompts, or external memory rather than model weights. Learning to Self-Evolve defines test-time inter-episode self-evolution as repeated context rewriting, formalized by $c_{t+1}=f_\psi(c_t,S_t)$, where $S_t$ is a structured summary of prior performance; at inference time, a tree-guided evolution loop composes many such edits while a fixed action model solves the task (Chen et al., 19 Mar 2026). Steve-Evolving likewise emphasizes non-parametric self-evolution: it keeps planner parameters fixed and instead anchors each embodied subgoal attempt into a structured tuple $$e_t=\langle s_{pre}^{(t)}, a_t, \mathcal D(s_t,a_t), s_{post}^{(t)}\rangle$$, distills successful trajectories into skills and repeated failures into executable guardrails, and injects those artifacts back into future planning (Xie et al., 13 Mar 2026). SEC-Depth also uses the model’s own history as supervision, but in a training-time contrastive form: historical latency models serve as “past selves,” and a self-evolution contrastive loss pushes current adverse-weather predictions toward clean predictions while separating them from weaker historical degraded predictions (Cao et al., 19 Nov 2025).

A second family uses self-generated critique, revision, or synthetic supervision. SELF first teaches meta-skills for self-feedback and self-refinement, then iteratively generates responses on unlabeled prompts, critiques them in natural language, revises them, filters the revisions, and fine-tunes on the resulting data (Lu et al., 2023). SEFT trains an adaptive reviser from preference-style revision pairs and then uses revised outputs on unlabeled prompts as pseudo-targets for ordinary supervised fine-tuning, combining internal evolution from the current policy and external evolution from a stronger generator (Chen et al., 2024). These methods treat revision quality, rather than reward-model optimization alone, as the proximate engine of policy improvement.

A third family co-evolves generator–solver or questioner–solver roles. INFUSER splits a pretrained checkpoint into a Generator $\pi_\phi$ and Solver $\pi_\theta$; the Generator drafts question–answer pairs from unstructured documents, while the Solver trains on them, and the Generator is rewarded not by difficulty alone but by an optimizer-aware influence score measuring whether a proposed question would improve the Solver on the target distribution (Chen et al., 8 Jun 2026). Video-Zero adapts the same co-evolutional idea to video: a Questioner discovers temporally localized evidence spans and generates evidence-grounded QA, while the Solver learns to answer and align its predictions with the supporting evidence, forming the loop $B$0 (Zhang et al., 14 May 2026). CurEvo adds curriculum guidance to video self-evolution by splitting generated supervision into perception, recognition, and reasoning dimensions and dynamically reallocating emphasis toward weaker dimensions (Zeng et al., 29 Apr 2026).

A fourth family evolves code, organization, or trajectories. Controlled Self-Evolution for algorithmic code optimization combines diversified planning initialization, feedback-guided genetic mutation and crossover, and hierarchical evolution memory to optimize correctness-constrained efficiency in open-ended program space (Hu et al., 12 Jan 2026). Self-EvolveRec extends open-ended code evolution to recommender systems by adding a User Simulator and Model Diagnosis Tool, then co-evolving the diagnostic tool with the recommender architecture itself (Kim et al., 13 Feb 2026). SE-Agent treats full LLM-agent trajectories as evolvable objects; prior reasoning traces are revised, recombined, and refined rather than merely ranked, which turns trajectory search into a population-based evolutionary process (Lin et al., 4 Aug 2025).

4. Learning signals and optimization objectives

The literature does not use a single canonical objective. Instead, self-evolution methods differ sharply in what they optimize and how they assign credit. Learning to Self-Evolve reduces multi-step self-evolution to a single-step RL objective over context edits and uses the improvement in downstream performance as reward, $B$1, explicitly avoiding a learned value baseline so that the edit-improvement signal is preserved (Chen et al., 19 Mar 2026). CurEvo, by contrast, uses evaluator-scored question quality $B$2, answer confidence $B$3, adaptive curriculum ratios $B$4, and weighted cross-entropy $B$5, with $B$6 updated from dimension-specific competence estimates $B$7 (Zeng et al., 29 Apr 2026).

Co-evolutionary methods often optimize usefulness rather than hardness. INFUSER defines an optimizer-aware influence score

$B$8

where $B$9 is the AdamW-preconditioned per-question Solver update direction, and then trains the Generator with DuGRPO, a dual-normalized variant of GRPO designed for continuous, noisy rewards (Chen et al., 8 Jun 2026). SELF-EMO uses a different signal entirely: valid emotion-prediction outputs receive a smoothed weighted IoU reward,

$C$0

and SELF-GRPO augments this with a group-level consistency reward over multiple rollouts from the same conversational prompt (Zhang et al., 20 Apr 2026).

In code and systems settings, direct environment-grounded utility is common. CSE evaluates candidate programs by the memory–time integral

$C$1

then defines reward as $C$2, with failed candidates assigned zero reward (Hu et al., 12 Jan 2026). Meta-Team models experience-driven MAS evolution abstractly as

$C$3

and instantiates $C$4 through L1 agent-level, L2 interaction-level, and L3 team-level scaffold updates rather than parameter learning (Hao et al., 28 May 2026). A common thread is therefore not a shared loss, but a shared commitment to converting self-generated experience into an evaluable update signal.

5. Domains of application and empirical evidence

Empirical studies show that self-evolution has been instantiated across software engineering, embodied control, video understanding, emotion modeling, recommendation, and prompt-based reasoning. In multi-agent reasoning, Meta-Team reports that collaborative experience organization outperforms both centralized and partitioned experience on long-horizon benchmarks; on Ansible, scores rise from 40.8 for no evolution to 49.8 for centralized experience and 53.9 for collaborative experience, and on ResearchRubrics from 49.5 to 52.9 and 55.8, respectively (Hao et al., 28 May 2026). In software issue resolution on SWE-bench Verified, SE-Agent reaches Pass@1/Pass@5 of 61.2/63.6 with Claude-3.7-Sonnet and is reported to deliver up to 55% relative improvement across five strong LLMs (Lin et al., 4 Aug 2025).

In open-world embodied planning, Steve-Evolving improves success rate on the Minecraft MCU suite across multiple backbones; with Gemini-3-pro, overall success rate is 53.37 for Steve-Evolving versus 47.63 for Optimus-1 and 42.67 for Jarvis-1, and the ablation shows that removing explicit knowledge injection causes the largest drop on hard groups (Xie et al., 13 Mar 2026). In robust depth estimation, SEC-Depth improves WeatherKITTI AbsRel from 0.120 in the baseline setting to 0.104 with the full CL + ID + $C$5 + $C$6 self-evolution configuration, while also improving zero-shot average AbsRel from 0.169 to 0.142 (Cao et al., 19 Nov 2025). In algorithmic code optimization, CSE reports on GPT-5 average ET/MP/MI of 67.70/65.62/67.47, exceeding AlphaEvolve and SE-Agent under the same 30-candidate budget (Hu et al., 12 Jan 2026).

Video understanding supplies a particularly clear testbed for grounded self-evolution. Video-Zero reports stronger evidence grounding than a full-video baseline, with Key Necessity of $C$7 versus $C$8 and Key Specificity of $C$9 versus $ODD_{Se} = ODD_{Si} \cup ODD_e$0, alongside gains across 13 benchmarks spanning temporal grounding, long-video understanding, and reasoning (Zhang et al., 14 May 2026). CurEvo shows that curriculum-guided self-evolution improves all seven tested video backbones on four VideoQA benchmarks; for example, Video-LLaVA(7B) on ActivityNet-QA rises from 45.36 to 48.72, and LLaVA-OneVision(7B) on NExT-QA from 66.07 to 70.12 (Zeng et al., 29 Apr 2026).

Self-evolution has also been used for alignment-style post-training and human-centric dialogue. SEFT, which fine-tunes on revised responses generated from unlabeled prompts, reaches 16.6 LC Win Rate, 13.7 Win Rate, and 7.47 on MT-Bench when augmented with 60K additional unlabeled prompts, outperforming the listed open baselines in AlpacaEval 2.0 LC Win Rate (Chen et al., 2024). SELF-EMO reports state-of-the-art ERC gains, improving accuracy by $ODD_{Se} = ODD_{Si} \cup ODD_e$1 on Qwen3-4B and $ODD_{Se} = ODD_{Si} \cup ODD_e$2 on Qwen3-8B on IEMOCAP, MELD, and EmoryNLP through its closed-loop recognition–self-emotion–expression pipeline (Zhang et al., 20 Apr 2026). In recommender systems, Self-EvolveRec consistently improves both ranking metrics and simulated user satisfaction over NAS and scalar-only LLM evolution, for instance raising SASRec on CDs from NDCG@5/HR@5 of 0.3610/0.4870 in DeepEvolve to 0.3865/0.5274 (Kim et al., 13 Feb 2026).

6. Limitations, misconceptions, and open problems

A persistent limitation is that many influential formulations are still architectural or conceptual rather than fully executable end-to-end methods. Self-evolving embodied AI explicitly presents a high-level framework rather than pseudocode, explicit training objectives, or formal evolution dynamics, and the ODD-based self-evolution paper is likewise a conceptual architecture centered on ODD extension, a computing warehouse, and sandbox validation rather than a complete synthesis algorithm (Feng et al., 4 Feb 2026, Weyns et al., 2023). This leaves the integration problem—how to coordinate multiple evolving modules stably and efficiently—substantially open.

Another recurring issue is stability. Self-evolving software agents report current limits in behavioural inheritance, robustness, and stability as environmental complexity increases (Robol et al., 29 Apr 2026). The LLM self-evolution survey identifies limited iterative gains, synthetic-data degradation, model collapse, catastrophic forgetting, and the broader stability–plasticity dilemma as unresolved research problems (Tao et al., 2024). Video-Zero observes that pseudo-label accuracy drops across later Questioner iterations as the generated curriculum becomes harder, which suggests a noise-accumulation ceiling in annotation-free co-evolution (Zhang et al., 14 May 2026).

Several methods are bottlenecked by the quality of evaluators, critics, or anchors. CurEvo reports that when the evaluator is weaker than the base model, gains shrink or become negative (Zeng et al., 29 Apr 2026). SEFT removes annotation requirements for policy-optimization prompts, but still requires a capable reviser trained from preference-style revision data and uses a critic model to assign revision labels (Chen et al., 2024). INFUSER relies on a small dev set as an anchor for credit assignment and shows that gains are strongest when the dev anchor and document pool are well aligned with the target domains (Chen et al., 8 Jun 2026). These results complicate any claim that self-evolution universally eliminates external supervision.

Safety, controllability, and accountability remain central open directions. The embodied-AI framework explicitly highlights controllable self-evolution, trustworthy self-evolution, and swarm self-evolution as major future directions, pointing to hierarchical control, interpretable feedback, stability margins or adaptation budgets, auditability, rollback mechanisms, and alignment under limited supervision (Feng et al., 4 Feb 2026). The ODD-based formulation similarly emphasizes uncertainty in ODD boundaries, novelty detection, compatibility checking, and sandbox-based evidence before enactment (Weyns et al., 2023). A common misunderstanding is therefore that self-evolution is intrinsically autonomous and self-justifying. The literature instead repeatedly treats verification, filtering, evaluator design, and deployment control as indispensable parts of the method.

Taken together, these works portray self-evolution method as a broad research program: systems are no longer restricted to fixed tasks, fixed representations, or fixed feedback channels, but are allowed to modify some part of their own operational substrate from experience. The specific object of evolution—context, memory, code, diagnostics, organization, embodiment, or trajectory—varies by domain, yet the underlying ambition is consistent: continual improvement through self-generated evidence, selective retention, and repeated re-entry into the control loop.