The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

1.1 Scaling Burdens in Model Development

• Model development is bottlenecked by resource-intensive training, costly experience generation and evaluation, and recurring post-deployment adaptation requiring repeated human diagnosis, revision, and regression testing.Examples include 9,240 expert-hours for GDPval, 3.2 million long-reasoning solutions in AIMO-2, and roughly ten engineer-hours to diagnose one infrastructure regression.Source: §1.1 · p. 3

1.2 From Development Burden to RSI

• RSI distinguishes autonomy, efficiency, and innovation by expanding AI responsibility from executing updates to identifying limitations, extracting experience, proposing changes, and validating successors.It targets both task performance and the mechanisms through which later improvements are discovered and implemented.Source: §1.2 · pp. 3–4

1.3 Challenges for RSI

• RSI faces three challenges: safely inheriting useful changes, attributing autonomy to AI-controlled decisions, and verifying gains without evaluator exploitation or increased search effort.Examples include performance degradation in 14% of Gödel Agent trials, externally fixed archive rules in Darwin Gödel Machine, and evaluator attacks addressed by frozen epochs and independent anchors.Source: §1.3 · p. 4

1.4 An Autonomy-Centered Framework

• The autonomy-centered framework ranks RSI from L1 execution autonomy through L2 strategy, L3 experience acquisition, L4 environment adaptation, and L5 recursive inheritance autonomy.Each level identifies what the AI internalizes, what persists across rounds, and which decisions remain externally controlled.Source: §1.4 · pp. 4–5

1.5 Application Domains

• Application analysis compares science, embodied intelligence, software engineering, and healthcare by feedback cost, reliability, retention, reuse, and remaining external control.These domains span uncertain experimental attribution, costly physical interaction, executable but incomplete tests, and high-stakes outcomes requiring expert oversight.Source: §1.5 · pp. 5–6

1.6 Industrial Evidence

• Industrial evidence complements academic literature by showing how evaluation pipelines, data flywheels, agent harnesses, automated experimentation, and deployment feedback operate under computational, feedback-quality, and human-involvement constraints.The survey characterizes demonstrated industrial responsibilities alongside more ambitious recursive-improvement visions; Figure 16 and Table 12 organize this landscape.Source: §1.6 · p. 6

1.7 Differences from Existing Surveys

• Compared with existing surveys, this work treats the complete improvement loop as its unit of analysis, defines autonomy by transferred responsibilities, and separates structural recursion from performance.It also compares mechanisms across science, embodied intelligence, software engineering, healthcare, and industrial settings with differing feedback and validation conditions.Source: §1.7 · pp. 6–7

2.1 Uneven Capability Progress in Modern Foundation Models

• By 2026, capability frontiers are uneven: advanced mathematics and graduate-level science reach HCI values of 86.4 and 85.8, versus 52.6 for software engineering and 39.9 for tool agents.Interactive capabilities retain larger gaps despite recent gains, while the illustrative post-2026 extension assigns larger improvements to domains with more unclosed headroom.Source: §2.1 · p. 9
• Software engineering and tool use are especially relevant RSI settings because their persistent updates require environment construction, trajectory collection, failure diagnosis, training or harness revision, and regression testing [25] [62] [63].Interactive workflows require planning, state tracking, tool selection, result interpretation, and action revision, so errors propagate across complete trajectories.Source: §2.1 · p. 9

2.2.1 Anatomy of an Improvement Loop

• An improvement loop uses experience to propose a modification, evaluates it under an acceptance rule, retains accepted changes in system state, and begins the next round from that updated state.Its anatomy tracks the AI system, state, experience, target, improver, strategy, verifier, improvement, and successor, while asking where the loop closes, what persists, and which decisions remain external.Source: §2.2.1 · p. 10

2.2.2 Definition of Recursive Self-Improvement

• RSI is the capability to transform experience and feedback into persistent self-changes that alter how subsequent self-improvements are generated, evaluated, selected, or consolidated.The improved system re-enters later interaction rounds with a changed capability state, allowing its own improvement capacity to participate in an ongoing recursive process.Source: §2.2.1 · p. 10, §2.2.2 · p. 11

2.2 Conceptual Foundations of Recursive Self-Improvement

• The section distinguishes RSI from predecessors and related approaches by requiring an autonomous closed loop in which accepted changes can affect the improvement process itself.Self-modifying programs, the Gödel Machine, AutoML, meta-learning, continual learning, open-ended learning, and agentic systems address overlapping components but do not alone establish this criterion [14] [65].Source: §2.2 · pp. 9–10, §2.2.2 · p. 11

2.2.3 Relation to Neighboring Paradigms

• Continual learning, AutoML, and agentic AI approach RSI only when validated improvements persist across rounds and revise how later adaptations, searches, evaluations, or selections are performed.Their usual objectives, update rules, search spaces, budgets, harnesses, tools, stopping rules, and acceptance criteria remain fixed by designers or external actors [28] [43].Source: §2.2.3 · pp. 11–12, Table 1 · p. 12

2.3 Scope of Evidence

• The survey includes peer-reviewed research alongside technical reports, engineering blogs, repositories, model documentation, and public benchmarks because contemporary RSI mechanisms often emerge first in industrial systems.This scope captures retained artifacts, evaluator and tool interactions, and whether improvement mechanisms continue beyond a single experiment while distinguishing demonstrated from inferred mechanisms.Source: §2.3 · p. 12

3 RSI Across Autonomy Levels

• The RSI hierarchy progresses from in-session refinement to persistent adaptation and ultimately recursive improvement of the mechanisms that produce future improvements.It evaluates each level by where the loop closes, what improvement persists, and which decisions remain under human control.Source: §3 · p. 12

3.1 B0: In-Task AI Improvement

• B0 improves current outputs but retains no persistent system change, so experience and correction signals cannot accumulate across independent tasks.Its procedures and evaluation rules remain human-defined, and repeated self-feedback or iteration may fail to correct errors under fixed evaluation.Source: §3.1 · pp. 12, 14

3.2.7 Evaluating L1 Execution Autonomy

• L1’s main risks are unreliable execution outside predefined procedures and unsafe persistence, because erroneous artifacts can propagate into subsequent improvement rounds.Evaluation therefore requires both execution reliability and persistence safety before retained outputs enter downstream workflows.Source: §3.2 · p. 15, §3.2.7 · p. 18

3.2 L1: Autonomy over Improvement Execution

• L1 systems execute human-defined improvement procedures and retain validated artifacts for reuse across later tasks and development stages.This capability spans data curation, synthetic supervision, training, infrastructure, evaluation, deployment, and application engineering, but the improvement procedures remain externally specified.Source: §3.2 · pp. 14–15, §3.2.1 · p. 16, §3.2.2 · p. 16

3.3 L2: Autonomy over Improvement Strategies

• At L2, the system uses evaluation feedback and experimental evidence to diagnose weaknesses, select the next intervention, test it, and retain accepted revisions.Humans still specify the objective, task boundary, and acceptance criteria; L2 transfers responsibility for deciding which improvement experiment to attempt next.Source: §3.3 · p. 18, Figure 4 · p. 19

3.3.1 Prompt Search

• In prompt search, L2 systems revise textual or multimodal controls from execution traces and evaluation results, using targeted feedback-driven search rather than merely applying prescribed edits.GEPA, for example, attributes failures to modules and preserves complementary prompt variants, while MPO and C-Evolve explore broader candidate revisions, [83].Source: §3.3.1 · p. 19

3.3.2 Agent and Harness Search

• Agent and harness search expands L2 from prompts to executable agent designs, allowing revisions to change reasoning, tool use, workflows, and behavior across entire tasks.Examples include Agent Optimizer and ADAS,, while AFlow and AgentSquare search executable graphs or modular designs through tree search, evolution, and recombination,.Source: §3.3.2 · p. 20
• Agent and harness search selects among alternative agent designs, but candidates remain bounded by externally maintained performance evaluations.This exceeds local parameter tuning without granting authority to redefine successful performance.Source: §3.3.2 · p. 20

3.3.3 Model and Training Search

• Model and training search spans configurations, architectures, training rules, and implementations, using empirical results to allocate experiments and retain changes that improve protected evaluations [11] [90].Representative L2 systems search these improvement objects while objectives, evaluation criteria, or acceptance rules remain externally specified.Source: §3.3.3 · p. 20, Table 4 · p. 21

3.3.4 Evaluating L2 Strategy Autonomy

• L2 autonomy can fail to produce durable improvement when search is inefficient, exploits evaluator weaknesses, or yields gains that disappear across iterations.Autonomy breadth and improvement-process quality are separate dimensions; broader intervention spaces help only when feedback discriminates candidates and search explores efficiently.Source: §3.3.4 · p. 21
• Adaptive evaluation risks benchmark overfitting, seed cherry-picking, shortcut discovery, test-label extraction, and confusing extra search compute with algorithmic progress.Repeated benchmark queries turn evaluation into part of the optimization surface, so company-reported feasibility evidence should not be treated as independently replicated results.Source: §3.3.4 · p. 21
• L2 shifts human effort from proposing individual improvements to constraining the search while retaining external objectives and acceptance criteria.The improver selects among agent designs, but promotion remains governed by an externally maintained evaluation process; L2 cannot redefine success.Source: §3.3.2 · p. 20, §3.3.4 · p. 21

3.4.1 From Industrial Automation to Experience Autonomy

• Automated data operations do not establish experience autonomy unless learner feedback changes subsequent acquisition decisions through a closed dependence on persistent learning.The distinction applies across data preparation, labeling, selection, construction, mixture, orchestration, and diagnosis; fixed algorithms can still adapt when conditioned on changing learner feedback.Source: §3.4.1 · p. 22

3.4.2 Adaptive Task Generation and Self-Play

• Adaptive task generation and self-play make future experience responsive to learner-dependent signals, combining difficulty estimation with executable or formal checks when available.Across SSP, AZR, R-Zero, STP, PSV, and VisPlay, feedback can be grounded in outcomes or proofs or estimated through consistency; these signals trade off verification scope, cost, and proxy reliability.Source: §3.4.2 · pp. 23–25

3.4.3 Autonomous Practice through Environment Interaction

• Autonomous environmental practice links capability assessment to later objectives through persistent resources such as reusable skills, reward-model judgments, or evolving curriculum memory.VOYAGER [84], SIMA 2 [24], and SEAgent illustrate the pattern, but unreliable skills, reward judgments, or guidebook entries can misdirect later learning.Source: §3.4.3 · pp. 25–26

3.4.4 Evaluating L3 Experience Autonomy

• Evaluating L3 requires showing both learner-conditioned acquisition and resulting benefit across rounds, while tracking experience validity, diversity, transfer, and corruption risks.Controls should compare adaptive acquisition with learner-independent schedules under matched data, environment, compute, and interaction budgets; protected evaluation must remain separate from feedback-driven learning.Source: §3.4.4 · p. 26

3.4 L3: Autonomy over Future Learning Experience

• L3 autonomously uses learner state, failures, or learning history to choose future experience, and persistent updates feed back into later acquisition decisions.The loop is observe learner state, choose or acquire experience, update persistent state, reshape future experience, and repeat; human-designed objectives and evaluators may remain.Source: §3.4 · p. 22, §3.4.4 · p. 26

3.5.1 Trajectory distillation

• Trajectory distillation converts histories into reusable textual memories, structured records, procedural skills, or executable artifacts that condition later tasks.Methods range from evolving notes and strategy graphs to skill libraries, callable tools, and sandboxed code, with evaluation spanning sequential streams, within-run behavior, and cost.Source: §3.5.1 · pp. 27–29

3.5.2 Iterative revision of the agent system

• Iterative revision changes agent skills, rubrics, harnesses, or weights, using decoupled audits, fixed executors, score-gated rollback, or self-evaluated experience channels.Harness Updating Is Not Harness Benefit [151] finds that updating ability is largely independent of base-model capability, whereas exploiting updated harnesses fails when artifacts are not activated or followed.Source: §3.5.2 · pp. 29–30

3.5.3 Selective retention and deployment of updates

• Selective retention governs admission, maintenance, and release: candidates require validation, libraries need drift management, and production fixes remain subject to engineering review.The Library Drift study [152] shows that unbounded accumulation degrades retrieval and that overly aggressive retirement performs worse than leaving the library unguided; filtering edits improves performance [153].Source: §3.5.3 · pp. 30–31

3.5.4 Evaluating L4 Deployment Autonomy

• L4 evaluation must distinguish deployment autonomy from adaptation quality by tracing each retained change to its evidence, later use, and effects under time and distribution shifts.Persistent failures can reflect wrong lessons, invalid scope, non-invocation, or loss of earlier capabilities [151] [152] [153], while humans retain objectives, boundaries, protected acceptance criteria, and final release authority.Source: §3.5.4 · p. 31

3.5 L4: Autonomy in Deployment and Environmental Adaptation

• L4 shifts autonomy from selecting experience to deciding which consequences persist, creating a loop from deployment interaction through adaptation, validation, reuse, and new feedback.Persistent changes shape later behavior and feedback, but consequential releases remain externally governed.Source: §3.5 · pp. 26–27, §3.5.4 · p. 31

3.6 L5: From Environmental Adaptation to Meta-Improvement

• L5 begins when AI revises and reuses the mechanism governing later improvement, while humans retain the mission, protected evaluation, editable components, and resource authority.The inherited mechanism may be an improver, evaluator, search policy, or research procedure; structural L5 requires persistence and later control, whereas effective L5 additionally requires better successors under comparable budgets.Source: §3.6 · pp. 31–32

3.6.1 Improving the Search Procedure

• STOP [15] selected a fourth-generation improver that outperformed its seed on all five transfer tasks, while weaker models regressed and some programs exploited budgets or evaluation bugs.Gödel Agent enabled recursive update-logic revision but saw 14 of 100 MGSM trials fall below the initial policy; HyperAgents [93] supported transfer without a significant final advantage for transferred initialization.Source: §3.6.1 · pp. 32–33

3.6.2 Improving Successor Evaluation

• RQGM [21] achieved a 71.7% held-out Polyglot pass rate versus 69.9% for HGM-H, using frozen-epoch evaluators and independent ground-truth anchors to govern replacement.The anchor, schedule, and orchestration remain externally fixed, and the stability argument applies within each frozen epoch rather than unrestricted objective changes.Source: §3.6.2 · p. 33

3.6.3 Revising Research Goals and Policies

• A-Evolve-Training [17] revised its research policy when development scores diverged from external gains, reaching 0.86 versus 0.87 for the top human submission under a fixed worker substrate and objective.Its inherited policy retained recipes, promoted or retired search directions, and unsuccessful approaches; autonomous revision of the objective was not demonstrated.Source: §3.6.3 · p. 34

3.6.4 Industrial Practice

• These systems automate experiment execution within researcher-designed workflows; under private scoring and a fixed cost budget, Weco reports seven accepted improvements over 100 unattended steps, but no statistically significant efficiency advantage.The results transferred to external tasks, while private evaluation, task families, and cost constraints remained human-designed.Source: §3.6.4 · p. 34

3.6.5 Evaluating L5 Recursive Improvement

• L5 evaluation should audit the revised artifact, its motivating evidence, later invocation, matched-budget effectiveness, transfer, retention, resource use, protected tasks, and stopping under cost or risk.SEA-Eval [156] and SEAGym [157] support trajectory-based assessment, while proposed private and adaptive-frontier tests target discovery, forgetting, reward hacking, and fuller autonomy.Source: §3.6.5 · pp. 34–35
• Accepted changes to improvers, evaluators, or research policies can govern subsequent rounds and sometimes transfer across tasks, but reliable accumulation across generations remains open [15] [21] [93].The practical objective is validated improvement per unit of computation and human effort under strong models, infrastructure, and independent evaluation.Source: §3.6.5 · p. 35

3.7 Cross-Level Synthesis

• The autonomy hierarchy progressively internalizes persistence, strategy selection, learning-agenda control, deployment adaptation, and finally recursive inheritance, but higher autonomy does not imply better improvement.Lower and intermediate levels have broader support, whereas L3–L4 are domain-dependent and end-to-end L5 evidence remains concentrated in bounded prototypes and emerging systems.Source: §3.7 · p. 35

4.1 S1: RSI for Science

• Scientific RSI must jointly evolve hypothesis generation, experimental execution, and evidence interpretation because discovery is open-ended and feedback is sparse, delayed, ambiguous, and validity-dependent.A scientific RSI loop generates hypotheses and experiments, evaluates them computationally or physically, and uses evidence or falsification to diagnose the system; safe inheritance requires provenance, scope, and uncertainty.Source: §4.1 · pp. 36–37, §4.1.4 · p. 38

4.1.4 Toward Recursively Improving Science Agents

• Current AI4Sci systems reach L2 as the strongest frontier, with limited L3 behavior, while end-to-end L4 adaptation and L5 improver evolution remain largely unexplored.Persistent updates occur through retraining and memory accumulation [159] [160], while systems autonomously select or apply updates to weights, tools, skills, workflows, and scaffolds [162] [163] [171]; adaptive exploration and cross-task reuse provide some L3 evidence [166] [172].Source: §4.1.4 · p. 38

4.2.5 Toward Recursively Improving Embodied Agents

• Embodied systems evolve future experience, skills, policies, world models, and evaluators, yet their objectives and update procedures remain fixed and physical inheritance requires attribution, safety, and rollback.Environment generators target capability-frontier tasks, harnesses assemble capabilities, policies act, and world models or evaluators interpret outcomes; validated feedback can update these components in subsequent cycles.Source: §4.2 · p. 38, §4.2.5 · p. 40, §4.2.4 · p. 40
• Embodied RSI has advanced from policy and harness updates toward learner-conditioned curricula and simulation-based agent–environment co-evolution, but no safe, persistent, autonomous real-world loop integrates these capabilities.Systems demonstrate L1–L4 capabilities through policy updates [182] [183] [189], skill and harness updates [111] [179] [180], learner-conditioned experience and skill retention [84] [178], and mostly simulated co-evolution [65] [174] [175] [177]; ENPIRE [181] approaches bounded L5 while retaining externally specified interfaces, evaluators, and safety constraints.Source: §4.2.5 · p. 40

4.3.4 Toward Recursively Improving Software-Engineering Agents

• Software-engineering RSI is organized as Development Challenge → Agent-Level Modification → Repository-Grounded Trial → Regression-Aware Selection → Versioned Inheritance, enabling persistent and reversible agent updates.Repository feedback evaluates modifications on current and held-out tasks before accepted versions are inherited for future development and self-improvement.Source: §4.3.4 · p. 42
• Software-engineering systems establish L2 as the strongest level, show emerging L3 and bounded L5 behavior, but lack L4 environment adaptation and full L5 improvement of the improver.L1 cross-task experience and skill learning appear [60] [191]; L2 harness, tool, and workflow changes occur under fixed objectives and evaluators [23] [63] [194]; L3 task generation is emerging [197]; SICA and DGM show bounded L5 characteristics [190], while HGM adds lineage-based meta-selection without adapting selection.Source: §4.3.4 · p. 42

4.4 S4: RSI for Healthcare

• Healthcare is introduced as an RSI scenario spanning diagnosis, medical imaging, and treatment planning, where clinical feedback drives persistent reuse of updated knowledge, reasoning strategies, and workflows.Feedback may come from diagnostic results, clinician corrections, or clinical guidelines, with improvements retained in memory, skill libraries, or model parameters.Source: §4.4 · p. 42

4.4.1 Evolving Clinical Memory and Knowledge

• Clinical memory systems convert completed cases into persistent knowledge, while newer methods distill reusable diagnostic primitives, recalibrate memory applicability, and track both clinical lessons and tool reliability.Most evidence still relies on simulated or retrospective cases with externally supplied labels, and real-world validation remains lacking.Source: §4.4.1 · pp. 42–43

4.4.2 Evolving Clinical Reasoning Strategies

• Clinical reasoning evolves through feedback-driven diagnostic policies, multi-agent coordination weights and retrieval scope, and practice-grounded skill repertoires that can be internally reinforced.EvoClinician revises question and examination selection using clinical yield and resource cost, while EvoMDT updates specialist-opinion reconciliation and PsychAgent internalizes selected counseling skills.Source: §4.4.2 · p. 43

4.4.3 Evolving Healthcare Tools and Workflows

• Healthcare tools and workflows evolve from verified execution trajectories and clinician corrections into composite tools, procedural skills, reusable safeguards, analytical procedures, and reversible workflows.MACRO and SkeMex expand the action space according to demonstrated utility, while TissueLab and HealthFlow construct and govern reusable procedures that require validation before cross-case inheritance.Source: §4.4.3 · p. 43

4.4.4 Toward Recursively Improving Healthcare Agents

• L1 persistent inheritance is established, while L2 systems autonomously update prompts, memories, policies, tools, workflows, and coordination under externally fixed objectives and gates; end-to-end L4 and L5 remain largely unexplored.L3-like elements and simulated L4 adaptation exist, but patient populations, task distributions, admission criteria, clinical realism, and evaluation remain externally determined.Source: §4.4.4 · p. 44

4.4 S4: RSI for Healthcare

• Healthcare RSI requires reliable longitudinal outcome attribution, safe feedback acquisition without unrestricted exploration, selective transfer across settings, and validated, traceable, monitored, and reversible capability commitments.These requirements address unsafe clinical trial-and-error, delayed and confounded feedback, and population- or institution-dependent improvements.Source: §4.4 · p. 42, §4.4.4 · p. 44

5.1 Theseus: Environment–Data–Model Co-Evolution for RSI Intelligence

• 5.1 Theseus operationalizes environment–data–model co-evolution by reconstructing environments, generating targeted training data, improving task models, and iterating environments for further learning.Across eight model–harness configurations, clean workspaces raised pass rates by 21.7–51.6 percentage points, while reconstructed environments raised aggregate rubric scores by 18.65–39.67 percentage points across five fixed pairings, with some scope-creep regressions.Source: Figure 10 · p. 45, §5.1 · p. 45

5.2 Lark: Building the Data Foundation for Reliable Enterprise-Level RSI

• 5.2 Lark combines enterprise knowledge graphs with automated evaluation and failure attribution, increasing human-rated usability from 52% to 65% and automated-evaluation usability from 47% to 56% over RAG.Its evaluator achieved approximately 84% agreement with human judgments, while humans retain responsibility for standards, critical samples, and important changes.Source: §5.2 · p. 46

5.3 Humanlaya: Delivery-Driven RSI for Data Quality Assurance

• 5.3 Humanlaya’s feedback-driven updates reduced key-defect packages from 9.0% to 3.7% and average human handling time from 48 to 27 minutes across 600 held-out task packages.Its outer loop analyzes review, customer, and usage evidence, proposes versioned changes to prompts, examples, instructions, and skills, then validates and human-approves updates before reuse.Source: §5.3 · p. 47

5.4 ModelBest: Zero-Human Industrial AI Engineering

• 5.4 ModelBest’s ForgeTrain matched Megatron-LM v0.15 on H100 in roughly 8 hours and reportedly surpassed it within 1.5–2.5 days, while ForgeStencil delivered 1.15–1.9× speedups over public state-of-the-art implementations.Reported FLOPs utilization rose from 40.1% to 44.1% for MiniCPM4-0.5B and from 47.0% to 50.9% for the 8B model; ForgeStencil’s median end-to-end speedup was 1.41×.Source: §5.4 · p. 48

5.5 Tencent Hunyuan: Experience-Driven Self-Improvement

• Hyra reports validation BPB of 0.9015 versus 0.9109 on nanochat AutoResearch, 76.4 s versus 77.5 s on nanoGPT Speedrun, and 0.771 versus 0.754 on SOL-ExecBench.Hyra can revise evaluation mechanisms by increasing granularity, strengthening comparison baselines, or closing reward-hacking loopholes before subsequent search continues.Source: §5.5 · p. 49

5.6 Agent-Native Research Lab: Verifiable Research Infrastructure for RSI

• 5.6 Agent-Native Research Lab’s Agent-Native Research Artifact raised prior-work question-answering accuracy from 72.4% to 93.7% and RE-Bench reproduction success from 57.4% to 64.4%.ARA preserves executable specifications, exploration branches, and raw evidence, enabling successors to inspect how results were obtained rather than inheriting only polished narratives.Source: §5.6 · p. 50

Cross-Component Diagnosis and Coordinated Improvement

• RSI systems need cross-component diagnosis that converts failures into testable intervention hypotheses and coordinates changes across data, models, harnesses, and environments.Selective freezing, targeted ablations, dependency records, and controlled comparisons can distinguish useful repairs from compensation for upstream defects.Source: §6 · p. 51

Learner-Conditioned Experience Acquisition and Reliable Learning Signals

• Experience acquisition must preserve validity, diversity, coverage, and durable learning value while adapting to the learner and allocating verification effort under matched budgets.AZR combines executable checks with learner-dependent task proposal, whereas R-Zero uses solver consistency as a difficulty proxy; neither correctness nor difficulty alone establishes learning value.Source: §6 · p. 51

Persistent-State Management and Reliable Reuse

• Persistent state requires validating artifact activation, faithful execution, applicability, dependencies, and downstream benefit as libraries evolve and artifacts transfer across successors.Library Drift shows that expanding skill libraries can degrade retrieval, while overly aggressive retirement can also be harmful [152]; HDSO’s paired evaluations provide an admission-test starting point [106].Source: §6 · p. 51

Governed Adaptation under Domain-Specific Feedback

• Adaptation must use domain-specific validation and explicit human authority because evidence differs across science, embodied intelligence, software, and healthcare, with errors potentially irreversible.Simulation, replay, and sandboxing can precede supervised or staged evaluation, while monitoring must test persistence across changing populations, tools, requirements, and environments.Source: §6 · pp. 51–52

Trustworthy Evolution of Improvement Mechanisms and Long-Horizon Evaluation

• Improvement mechanisms require independent assessment and long-horizon evaluation that distinguishes inherited structure from effective gains in producing better subsequent improvements.RQGM freezes its evaluator within an epoch and validates replacements against an independent anchor [21]; AIDE2 did not establish a statistically significant efficiency advantage for an evolved outer improver, while A-Evolve-Training illustrates policy inheritance within a fixed high-level objective [17].Source: §6 · p. 52

Resource-Aware Improvement, Human Collaboration, and Reproducible Infrastructure

• RSI evaluation must account for total resources, human collaboration, and reproducible cross-round evidence so that improvements are attributable, transferable, and cost-justified.Relevant measures include diagnosis, acquisition, search, verification, monitoring, review, rework, and time to validated capability; reusable artifacts should preserve parent state, evidence, configurations, decisions, provenance, and replay conditions.Source: §6 · pp. 52–53

Conclusion

• The roadmap spans improvement execution, strategy selection, experience acquisition, environmental adaptation, and recursive meta-improvement, using the improvement loop to track system changes, inherited capabilities, and human-controlled decisions.It is connected to scientific discovery, embodied intelligence, software engineering, and healthcare, where feedback availability, verification costs, and deployment constraints shape RSI progress.Source: §7 · p. 53
• Industry practices and preliminary empirical findings ground the roadmap in demonstrated capabilities and current limitations.Source: §7 · p. 53
• Long-horizon evaluation under explicit resource constraints and human oversight is needed to test whether RSI mechanisms produce transferable gains.Source: §7 · p. 53
• RSI ultimately aims to make each generation of AI improve future systems more reliably, efficiently, and conducive to novel discoveries.Source: §7 · p. 53

Landscape overview

• Figure 16 summarizes 491 surveyed papers across autonomy levels L1–L5, primary improvement targets, and sub-targets, with fractional weighting for papers covering multiple targets.The inner, middle, and outer rings represent autonomy levels, primary targets, and sub-targets, respectively; autonomy-level percentages are based on paper counts.Source: Figure 16 · p. 67, Appendix A · p. 67

Memory

• The taxonomy identifies memory targets spanning episodic experience, semantic knowledge, procedural strategies, memory organization, and memory operations.These components include interaction records and reflections, persistent knowledge stores, reusable procedures, schemas and indexes, and policies for writing, retrieval, updating, consolidation, prioritization, and forgetting.Source: Table 11 · p. 67, Appendix A · p. 68

Workflow, tools, and models

• Workflow, control, tools, and model targets cover agent graphs, planning, routing, verification loops, multi-agent protocols, scaffold code, tool interfaces and implementations, skills, weights, adapters, architectures, and inference policies.The taxonomy includes execution dependencies, task decomposition, scheduling, critic–revise and verify–retry procedures, agent coordination, callable capabilities, reusable behavioral modules, trainable parameters, and runtime configurations.Source: Appendix A · p. 68

Training, evaluation, environments, and co-evolution

• Training, evaluation, data, environment, output, and co-evolution targets include objectives and optimizers, schedules and search, judges and verifiers, experience data and simulators, programs and scientific artifacts, and jointly modified models and harnesses.The full taxonomy also covers curricula, feedback and credit assignment, world models, algorithms, multi-target harnesses, model–harness coevolution, and mechanisms that generate, evaluate, select, and apply changes across cycles.Source: Appendix A · p. 69

• Table 12 organizes the surveyed industrial systems by company archetype and improvement target.Source: Appendix B · p. 70

Scope and organization

• The snapshot covers 72 distinct companies and teams, with one row per product or representative work and organization by company archetype rather than geography.The RSI tag compactly maps entries to the B0–L5 scheme and does not claim that companies themselves use that label.Source: Company Archetypes in the Emerging RSI… · p. 70

Agentic optimization and verification

• Agentic optimization systems improve policies, skills, code, or verification through trace harnesses, instruction editing, code testing, verifier pipelines, and iterative evaluation.Examples include Agent Lightning v1.0, SkillOpt, agents code, ReVeal, Universal Verifier, and Meta Self-Taught Evaluator.Source: Company Archetypes in the Emerging RSI… · p. 71

Training and reasoning bootstrapping

• Training-oriented efforts use synthetic data, environments, self-generated reasoning, verifiers, self-play, challenge generation, voting, and reward signals to improve agent or reasoning policies.The table includes Alibaba/Qwen systems, AgentWorld, Qwen-Scope, DeepSeek R1, and Tencent AI Lab R-Zero.Source: Company Archetypes in the Emerging RSI… · p. 71

Persistent and long-horizon agents

• Persistent-agent and long-horizon systems target memory, skills, tool use, orchestration, planning, and swarm coordination for extended execution.Representative entries include MiniMax’s persistent cloud agents, Moonshot AI’s Kimi K3 and Agent Swarm, and Nous Research’s Hermes Agent.Source: Company Archetypes in the Emerging RSI… · p. 71

Computer use and model customization

• Computer-use and model-customization systems place improvement in virtual environments or user-directed training loops, spanning perception, planning, action, RL, policy weights, tool-use RL, and LLM-judge grading.The table identifies Zhipu AI/Z.AI AutoGLM, AgentRL, Thinking Machines Lab’s Tinker and Inkling, and related model-customization work.Source: Company Archetypes in the Emerging RSI… · p. 71

• The surveyed efforts make self-improvement, AI-for-AI, self-evolution, or recursive improvement central to their company or research thesis.This common orientation underlies the diverse systems and products listed in the section.Source: B. RSI-native / AI4AI companies · p. 72
• AI-chip and AI-research systems automate search, generation, execution, verification, evaluation, and iteration over designs, training recipes, code, hypotheses, or experiments.Examples include Ricursive Intelligence’s AI-chip co-design, Recursive Automated AI Research, Imbue Catalyst, Darwinian Evolver, Mirendil, AI Scientist, and Adaption Labs’ AutoScientist and Forge.Source: B. RSI-native / AI4AI companies · p. 72
• Several systems target software-agent evolution by updating skills, memory, code, harnesses, environments, or weights through feedback, refactoring, experience, evaluation, or search.This pattern appears in Evolvent AI Org, RSIBench, Terrarium, Frontis AI / Xianyuan OpenRSI, and related self-evolving-agent entries.Source: B. RSI-native / AI4AI companies · p. 72
• Other efforts evolve reusable code assets or research workflows by generating, testing, publishing, reusing, composing, or critiquing programs, capsules, hypotheses, and experiments.The listed examples include EvoMap, GEP, GeneBench, Endless Frontier BigBang-v1, ComfyResearch, and OPHIS.Source: B. RSI-native / AI4AI companies · p. 72
• Long-horizon and self-evolving research systems extend the target toward agents that accumulate experience, plan and review work, modify research or tool-use policies, and conduct experiments or code tasks.Examples include Chaoyan Intelligence’s TUMIX, R1-Code-Interpreter, Theseus Argus, SetupX, and related AI Scientist entries.Source: B. RSI-native / AI4AI companies · p. 72

• Periodic Labs develops laboratory AI that models science hypotheses, designs experiments, runs labs, and learns from experiment data.Source: C. Autonomous R&D &… · p. 73
• Edison Scientific’s Kosmos and Lila Sciences’ platform automate scientific discovery from hypotheses and code through experiments, post-training data design, lab execution, and real-time operation.Source: C. Autonomous R&D &… · p. 73
• Formal-mathematics systems from Axiom Math and Harmonic focus on conjecture, proof search, formal proof, verification, and translation.Source: C. Autonomous R&D &… · p. 73
• Autonomous model-R&D systems such as autoresearch, Prime Intellect, and Speedrun automate training recipes, optimizer and code hypotheses, GPU runs, and ablations.Source: C. Autonomous R&D &… · p. 73
• Research agents and related platforms automate research workflows involving ideas, tool use, experiments, coding, evaluation harnesses, benchmarks, multi-agent traces, and reports.Source: C. Autonomous R&D &… · p. 73

• Agent-improvement infrastructure spans feedback, skills, RL, evaluation, simulators, training data, and production-agent improvement.Source: D. Agent optimization, evaluation &… · p. 73
• Warp mines feedback and rewrites coding-agent skills, while LangSmith updates evaluator prompts from model feedback and evaluates at L1–L2.Source: D. Agent optimization, evaluation &… · p. 73
• Session mining identifies agent-behavior failure clusters, and memory systems measure, route, and reuse shared experience for L1–L2 improvement.Source: D. Agent optimization, evaluation &… · pp. 73–74
• Other systems optimize training corpora, distributed RL and experience infrastructure, interpretability-driven reward signals, and simulator-based evaluation with scenario generation, grading, and failure analysis.Source: D. Agent optimization, evaluation &… · p. 74

Embodied intelligence

• VLA R&D platforms connect models, data, training, evaluation, and real-robot deployment for embodied systems.Galaxea AI’s G0/G0.5 and GForge are listed as representative platforms spanning these components.Source: E. Embodied, world-model &… · p. 74
• Tactile-world-model efforts combine multimodal sensing, prediction, and manipulation policies for embodied control.TARS Robotics’ AWE and TacForeSight are identified as representative tactile world-model systems.Source: E. Embodied, world-model &… · p. 74

World models and embodied data

• World-model and embodied-data companies pursue post-training data synthesis, environment modeling, evaluation, and deployment-oriented model development.The listed examples include SYNWorld, OctoMind, and OctoSense, alongside data, world-model, training, and evaluation components.Source: E. Embodied, world-model &… · p. 75

Continual and test-time adaptation

• Continual and test-time adaptation systems use world codes, fast weights, LoRA, skill vectors, and context adaptation to model environments or transfer skills.MirrOS Code as Worlds and Spatial-TTT emphasize world-model-based test-time adaptation, while SHINE, LIFT, and PaST emphasize continual learning and skill transfer.Source: E. Embodied, world-model &… · p. 75

Self-evolving agents

• Collaborative self-evolving agents target shared state and organizational knowledge, representing a more advanced adaptation direction.Nexus is listed under collaborative self-evolving agents with an L3-L4 candidate harness.Source: E. Embodied, world-model &… · p. 75

F. Persistent-memory & personal-AI companies

• Cross-task memory, identity, and reusable skills form the main substrate for adaptation.Source: F. Persistent-memory & personal-AI… · p. 75
• Enterprise and local-agent efforts emphasize persistent knowledge, workspace memory, experience capture, and storage across web, files, and code.Source: F. Persistent-memory & personal-AI… · p. 75, E. Embodied, world-model &… · p. 75
• Agent-memory operating-system efforts provide long-term memory, skill retrieval, skill forging, and active-task support.Examples include EverMind, Raven, and EverMemOS.Source: F. Persistent-memory & personal-AI… · p. 75
• Personal-AI efforts explore identity models, LoRA-based memory modeling, personalization, and personal memory.Mindverse Second Me is listed within this direction.Source: F. Persistent-memory & personal-AI… · p. 75
• The review distinguishes task-local autonomy, increasing autonomy levels, adjacent infrastructure, plausible candidates, and explicit future targets, with some entries requiring stronger verification.Source: F. Persistent-memory & personal-AI… · p. 75