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

1.1 Scaling Burdens in Model Development

• Foundation-model training, experience generation and evaluation, and post-deployment adaptation remain costly, externally coordinated bottlenecks involving experimentation, feedback infrastructure, and repeated human-led diagnosis and testing.Examples include expert-intensive data curation, large post-training or synthetic-experience pipelines, and engineers manually diagnosing failures and running regression tests after deployment.Source: §1.1 · p. 3

1.2 From Development Burden to RSI

• RSI is an autonomous closed loop that identifies limitations, develops and validates improvements, then uses resulting capabilities to improve future improvement itself.A-Evolve-Training retains a research policy that redirects later experiments; across four rounds, its external score rose from 0.80 to 0.86, versus 0.87 for the top human submission. Ouroboros uses reviewed deployment evidence to revise coding-agent tools, context, prompts, or implementation, with tests and human review before deployment.Source: §1.2 · pp. 3–4

1.3 Challenges for RSI

• Credible RSI requires safe inheritance, attributable autonomy, and reliable verification, because persistent updates can regress, fixed search rules can mask autonomy, and evaluators can be exploited.In Gödel Agent, 14% of 100 MGSM optimization trials ended below the initial policy’s performance; the Darwin Gödel Machine’s SWE-bench subset score rose from 20% to 50% while archive maintenance and parent selection remained outside self-modification. Red Queen Gödel Machine freezes evaluators within each epoch and checks replacements against an independent ground-truth anchor.Source: §1.3 · p. 4

1.4 An Autonomy-Centered Framework

• The framework assigns increasing responsibility across execution, strategy choice, experience acquisition, environment adaptation, and recursive revision of improvement mechanisms.At earlier levels, humans retain the objective or evaluation criteria; later levels let systems choose experience, update persistent state through deployment interaction, or revise the mechanism governing future improvement.Source: §1.4 · p. 5

1.5 Application Domains

• Across science, embodied intelligence, software engineering, and healthcare, the survey compares what is updated and how feedback supports retention, from uncertain experiments and costly physical trials to executable tests and delayed clinical outcomes.These settings also differ in feedback attribution, repeatability, specification completeness, and the need for expert oversight.Source: §1.5 · p. 6

1.6 Industrial Evidence

• Industrial materials complement academic literature by showing how improvement responsibilities are handled within operating pipelines and how cost, feedback quality, and human involvement constrain deployed systems.The analysis distinguishes mechanisms demonstrated in production-oriented settings from more ambitious recursive-improvement visions that remain to be validated.Source: §1.6 · p. 6

1.7 Differences from Existing Surveys

• Unlike prior surveys organized by stage, update target, timing, or mechanism [26] [28] [29], this survey traces full-loop decisions and measures autonomy by responsibilities transferred from external designers to AI [31] [32].Its loop analysis follows what triggers improvement, who proposes and validates changes, what persists, and which later decisions use those changes.Source: §1.7 · pp. 6–7
• The survey separates structural from effective recursion and compares improvement mechanisms across operating conditions, asking what evidence supports recursive progress.Structural recursion means a revised improvement mechanism governs a later round; effective recursion means it produces stronger successors under comparable budgets and independent evaluation. The comparison spans science, embodied intelligence, software engineering, healthcare, and industrial practice, where feedback and external control differ.Source: §1.7 · p. 7

2.1 Uneven Capability Progress in Modern Foundation Models

• By 2026, advanced mathematics and graduate-level science reach HCI 86.4 and 85.8, versus software engineering 52.6, search and terminal agents 56.8, and tool agents 39.9.Interactive capabilities retain larger gaps, and gains in bounded or readily verified environments have not transferred uniformly to long, stateful workflows requiring planning, state tracking, tool use, verification, memory, and adaptation.Source: Figure 3 · p. 8, §2.1 · p. 9

2.2 Conceptual Foundations of Recursive Self-Improvement

• Analyzing RSI through the improvement loop enables comparison across mechanisms by asking what generates experience, what changes and persists, who controls updates, and whether one round shapes the next.Earlier traditions—including self-modifying programs, AutoML, meta-learning, continual learning, and agentic systems—address parts of the problem but do not alone establish RSI [14] [65].Source: §2.2 · pp. 9–10

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.RSI additionally requires sufficient autonomy and that accepted changes can affect how subsequent improvements are generated, evaluated, selected, or consolidated.Source: §2.2.1 · p. 10

2.2.2 Definition of Recursive Self-Improvement

• RSI is autonomous, persistent self-change from interaction-derived experience that can alter the mechanisms used to generate, evaluate, select, or consolidate later self-improvements.The changed system re-enters subsequent interaction and improvement rounds, allowing its capacity for improvement itself to participate in an ongoing recursive process.Source: §2.2.2 · p. 11

2.2.3 Relation to Neighboring Paradigms

• Continual learning, AutoML, and agentic AI automate parts of improvement, but approach RSI only when validated changes persist and can revise later improvement processes [28] [43].Their learning rules, search spaces, or task harnesses and acceptance criteria are generally fixed in advance; RSI instead asks whether an improvement loop forms around the system itself and how much authority becomes endogenous.Source: §2.2.3 · pp. 11–12

2.3 Scope of Evidence

• Because contemporary RSI spans academic research and industrial engineering, the survey includes concrete operational evidence from reports, engineering blogs, repositories, model releases, and benchmark leaderboards.This broader scope captures how loops operate and what persists across iterations while distinguishing demonstrated mechanisms from inferred ones.Source: §2.3 · p. 12

3.1 B0: In-Task AI Improvement

• At B0, systems refine current-task outputs but retain no persistent system changes, so experience cannot accumulate and more iterations under fixed evaluation may yield limited gains.Self-generated feedback can reinforce errors when verification is unreliable, while fixed verifiers may miss errors.Source: §3.1 · p. 14

3.2.7 Evaluating L1 Execution Autonomy

• L1 execution remains bounded by predefined procedures; out-of-scope conditions can cause failures, and retained errors can propagate into later workflows.Evaluation should assess both execution reliability and whether artifacts are validated before persistence and downstream reuse.Source: §3.2 · p. 15, §3.2.7 · p. 18

3.2.1 Data Level; 3.2.2 Training-Method Level

• At the data and training-method levels, L1 systems apply human-set criteria and workflows to curate or generate training data and to revise outputs or produce training signals.Procedures include data filtering and synthetic supervision generation, as well as rubric-guided revision and evaluation for subsequent optimization.Source: §3.2.1 · p. 16, §3.2.2 · p. 16

3.2.3 Training-Platform Level; 3.2.4 Evaluation-and-Safety Level; 3.2.5 Deployment-Optimization Level; 3.2.6 Application-System Level

• At the training-platform, evaluation-and-safety, deployment, and application levels, L1 agents execute predefined optimization or recovery, assessment, hardware adaptation, and software-integration procedures.Humans continue to define evaluation criteria, target platforms, product requirements, and development constraints.Source: §3.2.3 · p. 17, §3.2.4 · p. 17, §3.2.5 · p. 17

3.2 L1: Autonomy over Improvement Execution

• L1 executes human-defined improvement procedures at scale, retaining artifacts for later development workflows while humans specify the steps and validation rules.Meta’s Capacity Efficiency system encodes debugging expertise into reusable repair skills; the paper reports that it compresses hours of regression investigation into minutes.Source: §3.2 · p. 14, §3.2.7 · p. 18

3.3 L2: Autonomy over Improvement Strategies

• At L2, systems use evaluation evidence to choose and test the next improvement intervention, while humans retain control of objectives, task boundaries, and acceptance criteria.The search target can range from prompts to agents or model-training programs; autonomy concerns choosing what change to attempt, not unrestricted research control.Source: §3.3 · pp. 18–19

3.3.1 Prompt Search; 3.3.2 Agent and Harness Search

• Prompt search uses execution traces and evaluation feedback to select revisions, while agent and harness search can alter executable components or workflows across an entire task.Approaches include reflective prompt revision and multimodal prompt candidates, as well as agent archives, workflow graphs, and reusable modules.Source: §3.3.1 · p. 19, §3.3.2 · p. 20

3.3.3 Model and Training Search

• L2 systems search over configurations, model architectures, training rules, and implementations, while external objectives, validation criteria, or correctness tests govern acceptance.Search can use prior training evidence to allocate experiments, propose structures, modify learning procedures, or optimize implementations under protected checks.Source: §3.3.3 · pp. 20–21

3.3.4 Evaluating L2 Strategy Autonomy

• L2 shifts human effort from proposing each intervention to constraining the search; success criteria remain externally specified, and broader autonomy does not guarantee efficient or durable improvement.Adaptive benchmark queries can cause overfitting, confuse added search compute with algorithmic gains, or exploit evaluator blind spots.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: the defining evidence is a closed loop in which learner feedback changes later experience acquisition.Reviewed pipelines span six recurring functions, from data-source preparation and labeling to selection, construction, mixture and orchestration, and model-feedback diagnosis.Source: §3.4.1 · p. 22

3.4.2 Adaptive Task Generation and Self-Play

• Adaptive task-generation systems use changing learner signals to shape future tasks, while execution or formal checks can validate task properties without guaranteeing useful difficulty or learning gain.Consistency-based difficulty signals are proxies rather than correctness evidence; PSV’s targeted easy, medium, and hard problem difficulties substantially overlap, limiting curriculum control.Source: §3.4.2 · pp. 23–25

3.4.3 Autonomous Practice through Environment Interaction

• VOYAGER, SIMA 2’s ASKA setup, and SEAgent connect capability or interaction feedback to later practice through persistent skills, reward assessments, or an evolving software guidebook.The coupled loop’s reliability depends on the quality of skill construction, reward judgments, and stored guidebook information; SIMA 2’s fixed-task experiment alone does not establish task-selection autonomy.Source: §3.4.3 · pp. 25–26

3.4.4 Evaluating L3 Experience Autonomy

• L3 evaluation must test both learner-conditioned acquisition and its benefit, tracking experience validity, multi-round performance, and transfer while protecting evaluation data from adaptive use.Controls can freeze learner-state inputs or replace adaptive acquisition with learner-independent schedules, while matching data access and compute or interaction budgets.Source: §3.4.4 · p. 26

3.4 L3: Autonomy over Future Learning Experience

• L3 uses evolving capabilities, failures, or learning history to steer the next experience, with persistent learning feeding back into later acquisition decisions.This learner-conditioned loop may use a separate curriculum component or the agent’s policy, and persistent learning can reside in parameters or reusable external state.Source: §3.4 · p. 22, §3.4.4 · p. 26

3.5 L4: Autonomy in Deployment and Environmental Adaptation

• L4 uses deployment interaction to decide which experience-derived changes persist in agent components and shape later behavior, while objectives, access boundaries, evaluation, and consequential releases remain externally governed.The loop proposes adaptation, revises components, validates and retains accepted changes, then reuses them in subsequent tasks.Source: §3.5 · pp. 26–27, Figure 7 · p. 27

3.5.1 Trajectory distillation

• Trajectory distillation converts interaction histories into compact artifacts that persist across sessions and condition later tasks, including textual or structured experience memory, procedural skill libraries, and executable code.Trace2Skill [86] and PRACTICE [146] distill skills offline, whereas PANDO [25] acquires them during deployment.Source: §3.5.1 · pp. 27, 29

3.5.2 Iterative revision of the agent system

• L4 system revision can produce useful persistent updates without ensuring agents exploit them: updating ability is largely independent of base-model capability, while benefit varies non-monotonically.Failures arise when relevant artifacts are not activated or are activated but not faithfully followed; the analysis recommends investing capability in the task-solving agent rather than the evolver.Source: §3.5.2 · p. 30

3.5.3 Selective retention and deployment of updates

• L4 update governance requires validating proposed changes, maintaining libraries against drift, and gating consequential production releases.HDSO [106] tests candidate skills against an approved repository on increasingly large and independent tasks. Library Drift [152] finds that unbounded accumulation degrades retrieval and that overly aggressive retirement performs worse than unguided libraries; filtered edits also improve performance [153]. Tax AI routes fixes through engineering review, with humans retaining final acceptance.Source: §3.5.3 · pp. 30–31

3.5.4 Evaluating L4 Deployment Autonomy

• L4 evaluation must connect retained changes to their originating evidence, later use, and effects on new and previously solved tasks across time and distribution shifts.Persistent failures can reflect wrong lessons, out-of-scope application, failure to invoke updates, or loss of earlier capabilities; bounded task streams and incomplete or endogenous feedback leave effectiveness over time uncertain.Source: §3.5.4 · p. 31

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-model runs regressed on average.The selected improver governs the next round, but task distribution, utility, base model, and resource limits remain externally specified; some programs evaded soft budgets or exploited evaluation bugs. Gödel Agent also modifies its recursive update logic, but 14 of 100 MGSM trials ended below the initial policy.Source: §3.6.1 · pp. 32–33

3.6.2 Improving Successor Evaluation

• On held-out Polyglot coding tasks, RQGM [21] reports a 71.7% pass rate versus 69.9% for HGM-H, with lower search-token use.A frozen evaluator governs each epoch; challenger evaluators are checked against an independent ground-truth anchor, and scores tied to a replaced evaluator are discarded. The anchor and replacement schedule remain fixed.Source: §3.6.2 · p. 33

3.6.3 Revising Research Goals and Policies

• A-Evolve-Training [17] autonomously post-trained a 30B model across four rounds, reaching a final leaderboard score of 0.86 versus 0.87 for the top human submission.Its meta-agent revised a research policy from workers’ recipe changes, checkpoint evaluations, and failures; the worker substrate and constitution remained fixed, and the objective was not autonomously revised.Source: §3.6.3 · p. 34

3.6.4 Industrial Practice

• Automated Alignment Researchers generate and test mitigations within researcher-designed workflows; Weco reports seven accepted improvements in 100 unattended steps, yet its stronger outer-search test found no statistically significant efficiency advantage.Source: §3.6.4 · p. 34

3.6.5 Evaluating L5 Recursive Improvement

• A mechanism audit should trace each revision from motivating evidence to later invocation, then compare original and revised mechanisms from comparable agents under matched budgets, including evaluation costs.Keeping the revised mechanism fixed in transfer tests helps isolate what it learned about improvement; HyperAgents [93] uses this design to distinguish transferable agent-generation ability from original-task performance.Source: §3.6.5 · p. 34
• Proposed fuller-L5 evaluations combine adaptive frontier tasks with protected reference tasks, undisclosed-rule tasks, and checks that the system stops when expected benefit falls below cost or risk.These are proposed assessment requirements, not capabilities established by current RSI benchmarks.Source: §3.6.5 · pp. 34–35
• Accepted changes to improvers, evaluators, or research policies can govern later rounds, with some cross-task transfer [15] [21] [93]; reliable accumulation under comparable resources remains open [93].At L5, successors inherit a revised improvement procedure and its supporting evidence, while humans retain the overall objective, protected acceptance criteria, and resource authority.Source: §3.6.5 · p. 35

3.7 Cross-Level Synthesis

• Across B0–L5, each transition delegates another part of the improvement loop, from persistence and strategy choice to learning agendas, deployment adaptation, and recursive inheritance.The levels differ in internalized decisions, persistent state, and externally protected acceptance conditions.Source: §3.7 · p. 35
• Higher autonomy does not itself ensure better improvement: search can remain inefficient, feedback unreliable, and systems vulnerable to regression, evaluator exploitation, or poor transfer.Lower and intermediate levels have broader system support, while L3–L4 remain domain-dependent and end-to-end L5 evidence is concentrated in bounded prototypes and emerging systems.Source: §3.7 · p. 35

4.1 S1: RSI for Science

• Scientific RSI converts experimental outcomes into reusable agent changes, but evaluators and update-promotion rules generally remain fixed.Examples include reusable hypothesis-generation skills in HypoForge [158], validated tools in Test-Time Tool Evolution [159], transferred procedures in DrugSAGE [160], and scaffold, tool, search, and weight updates in SIA [161]. Improving a scientific artifact alone does not establish that the agent improved itself.Source: §4.1 · p. 36

4.2 S2: RSI for Embodied Intelligence

• Embodied RSI uses physical interaction and causal feedback to retain updates across interactions, but autonomous real-world validation and evolution of the improver remain unresolved.POET [65] co-evolves environments and agents, Voyager [84] stores successful behaviors as reusable skills, and World-VLA-Loop [162] improves a world model from policy failures and then optimizes the policy. Current loops often rely on simulation or externally specified rewards, evaluators, interfaces, and safety constraints; endogenous experience, difficult failure attribution, and costly, unsafe trials constrain real-world adaptation.Source: §4.2 · pp. 36–37

4.3 S3: RSI for Software Engineering

• Software-engineering RSI can execute, version, and roll back agent changes, enabling auditable inheritance, but remains bounded by a fixed outer improvement contract.SWE-Exp [60] reuses lessons from issue-resolution trajectories; Self-Harness [23] retains harness changes after regression testing; and Darwin Gödel Machine selects lineages by executable task performance. Repository tests remain incomplete proxies, and benchmarks, utility functions, sandboxes, and selection rules are externally defined.Source: §4.3 · p. 37

4.4 S4: RSI for Healthcare

• Clinical RSI reuses validated updates from longitudinal feedback, but evidence is mostly simulated or retrospective and real outcomes require careful oversight.Systems retain cases and diagnostic rules, revise evidence-acquisition processes, or compose verified trajectories into tools [163] [164] [165]. Updates should be provenance-aware, population-specific, prospectively monitored, reversible, and approved by experts and institutions; autonomous evolution of the clinical improver is not established.Source: §4.4 · p. 37

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

• Clean workspaces raised pass rates by 21.7–51.6 percentage points across eight configurations, while reconstructed environments raised aggregate rubric scores by 18.65–39.67 points across five fixed pairings.The studies used 30 workspace tasks; the reconstructed environment combined a Collection Map and Event Log. Some tasks showed scope-creep regressions, and a reusable file-verification module was trained as the co-evolution loop continues.Source: §5.1 · pp. 37–39, Table 9 · p. 38

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

• Lark’s automated data-quality evaluator achieved approximately 84% agreement with human judgments, and graph-based retrieval raised human-rated task usability from 52% to 65%.Automated evaluation and failure attribution guide concrete data corrections and later iterations; human reviewers retain responsibility for standards, critical samples, and important change approvals. Automated-evaluation usability with graph-based retrieval rose from 47% to 56% versus the RAG-based baseline.Source: §5.2 · pp. 39–40

5.3 Xiaohongshu: Dual-Timescale RSI for Recommendation

• On Discovery Feed, score_mean@16 rose from 2.211 to 2.366 and score_median@16 from 1.587 to 1.663; in-video-feed score_mean@16 rose from 1.537 to 1.739.These company-reported ranking-score improvements are preliminary and do not establish click-through or conversion gains. IMA updates persistent user memory from behavioral feedback, while reviewed hard cases support slower post-training consolidation.Source: §5.3 · pp. 40–41, Table 10 · p. 40

5.4 Humanlaya: Delivery-Driven RSI for Data Quality Assurance

• After four feedback-driven updates, Humanlaya’s key-defect package rate fell from 9.0% to 3.7%, and average human handling time fell from 48 to 27 minutes per task.The outer loop updates quality-system rules, prompts, examples, and skills using delivery feedback; candidate changes are tested on held-out task packages and reviewed by humans before promotion. Human oversight remains at critical stages.Source: §5.4 · pp. 41–42

5.5 ModelBest: Zero-Human Industrial AI Engineering

• ForgeTrain matched Megatron-LM v0.15 within roughly 8 hours and reportedly surpassed it within 1.5–2.5 days; ForgeStencil reported 1.15–1.9× speedups over public state-of-the-art implementations.ForgeTrain’s model 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 reported a median 1.41× end-to-end speedup. Forge Engineering combines within-project implementation optimization with cross-project reuse of engineering experience.Source: §5.5 · pp. 42–43

5.6 Tencent Hunyuan: Experience-Driven Self-Improvement

• Hyra reported validation BPB of 0.9015 on nanochat AutoResearch versus a cited previous best of 0.9109, and 0.771 on SOL-ExecBench [172] versus 0.754.Its Experience Bank retains solutions, artifacts, execution logs, scores, and evaluator feedback for later proposals; accumulated experience can also refine incomplete or exploitable evaluators. Hyra additionally reported 76.4 s on nanoGPT Speedrun versus 77.5 s.Source: §5.6 · pp. 43–44, Table 11 · p. 44

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

• 93.7% question-answering accuracy using ARA compared with 72.4% using conventional papers, while RE-Bench reproduction success rose from 57.4% to 64.4%.ARA preserves executable specifications, exploration branches, and raw evidence; the lab also gates claims on execution-trace or formal verification and reports a Chip-Bench Level-3 CPU score of 5.8416, 2.7% faster and 21.5% smaller than standard-agent baselines while passing all 97 directed tests and 350 random differential programs.Source: §5.7 · pp. 44–45

5.8 Frontis.AI: Enterprise Agent Evolution and Cross-Task Meta-Improvement

• Frontis.AI’s ME–WE–MA architecture separates human interaction, contextual memory, and release approval from execution by more than 300 specialist agents and agent construction, evaluation, and evolution.Its stated objective is to make improvement experience reusable across tasks rather than isolated within individual agents.Source: §5.8 · p. 45
• Shared evaluation infrastructure, bounded objectives, human-confirmed changes, and three validation gates support auditable, rollback-capable releases while preserving validated knowledge about improving agents.The gates assess regression safety, target achievement, and robustness in exceptional scenarios; approved updates return to production with versioned test reports.Source: §5.8 · pp. 45–46
• Frontis retains improvement episodes and transfers lessons to new tasks, adapting and revalidating them for receiving-task conditions rather than merely reusing task-performance knowledge.For example, prioritizing real-time evidence still requires identifying authoritative sources and checking their temporal validity.Source: §5.8 · pp. 45–46
• Across 76 comparable tasks, average scores rose from 5.25 to 5.87; evolution on unseen tasks was approximately 20% faster, based on company-reported evidence.Over two months, Frontis initiated 216 improvement tasks involving 117 specialist agents, and 63 completed at least one full evolution cycle; median machine-processing time was 27.8 minutes per completed cycle. The faster-evolution comparison used more than 100 tasks of accumulated meta-evolution experience and internal tests.Source: §5.8 · p. 46
• Frontis-MA1, a separately trained 35B model, increased its MLE-Bench Lite score from 39.39% to 60.61% under a fixed harness and reached 71.21% with additional search.The results concern machine-learning and deep-learning engineering; learning directly from enterprise-agent evolution trajectories remains future work.Source: §5.8 · p. 46

Cross-Component Diagnosis and Coordinated Improvement

• Because failures can arise across data, context, tools, models, evaluators, or environments, systems should test intervention hypotheses and coordinate changes across dependent components.Controlled comparisons, targeted ablations, selective freezing, and dependency records can distinguish useful repairs from compensation for upstream defects.Source: §6 · p. 46

Learner-Conditioned Experience Acquisition and Reliable Learning Signals

• Experience acquisition must assess validity, learner-dependent difficulty, and durable learning value, since correctness or estimated difficulty alone does not establish benefit.Mechanisms should preserve diversity and capability coverage, estimate value across rounds, and seek stronger external feedback when internal judgments are insufficient.Source: §6 · pp. 46–47

Persistent-State Management and Reliable Reuse

• Persistent artifacts require evidence of applicability and downstream benefit: retrieval can fail, agents may misuse retrieved skills, and library growth or over-retirement can harm improvement [152].Measure update quality, activation, faithful use, and later-task benefit separately, and revalidate artifacts as tasks, executors, and systems change; paired admission tests such as HDSO’s offer a starting point [106].Source: §6 · p. 47

Governed Adaptation under Domain-Specific Feedback

• Improvement policies must match validation to domain-specific feedback and risk, since evidence from science, embodied intelligence, software, and healthcare has different limits.Physical or clinical consequences cannot be undone by restoring a checkpoint, so pre-release validation and explicit human authority over consequential releases are essential.Source: §6 · p. 47

Trustworthy Evolution of Improvement Mechanisms

• Adaptive improvers need independently maintained acceptance criteria because changes to solvers, evaluators, or research policies can undermine score comparability and invite exploitation.RQGM freezes its evaluator within an epoch, validates replacements against an independent anchor, and revisits affected scores [21]; policy transfer beyond its original research setting remains untested [17].Source: §6 · pp. 47–48

Long-Horizon Evaluation of Inherited Improvement Capacity

• Structural evidence that an L5 mechanism is inherited and invoked does not establish effective improvement; AIDE2 found no statistically significant efficiency advantage from an evolved harness as the outer improver.Benchmarks should compare mechanisms from comparable starting points under matched total budgets, report full trajectories and costs, and test transfer on fresh tasks.Source: §6 · p. 48

Resource-Aware Improvement and Human Collaboration

• RSI efficiency must count diagnosis, acquisition, search, verification, monitoring, infrastructure, and human review—not just faster execution or longer unattended runs.Evaluate time and total cost to a validated capability target, alongside review and rework, and set explicit conditions for stopping, escalating to a human, or ending unproductive search; Meta’s workflow retains engineering review.Source: §6 · p. 48

Reproducible Infrastructure for Cross-Round Inheritance

• Reusable research artifacts should link each system state and proposed change to its evidence, evaluation, acceptance decision, and subsequent use.Versioned data, environments, and evaluators can support replay, but deterministic re-extraction and formal verification alone cannot establish that an experiment captures the intended capability.Source: §6 · p. 48

• The roadmap spans improvement execution, strategy selection, experience acquisition, environmental adaptation, and recursive meta-improvement.It analyzes what systems change, what successors inherit, and which decisions remain human-controlled.Source: §7 · p. 49
• Progress toward RSI across scientific discovery, embodied intelligence, software engineering, and healthcare depends on feedback availability, verification costs, and deployment constraints.Source: §7 · p. 49
• Industry practices and preliminary empirical findings ground the roadmap in demonstrated capabilities and current limitations, while long-horizon evaluation remains necessary to test transferable gains under explicit resource constraints and human oversight.The paper frames RSI's promise as making future improvement more reliable, efficient, and conducive to novel discoveries.Source: §7 · p. 49

Appendix A: RSI Landscape

• Figure 18 maps 491 surveyed papers across autonomy levels L1–L5, primary improvement targets, and sub-targets, with fractional weights for papers covering multiple targets.Autonomy-level percentages reflect paper counts; Table 12 details the target taxonomy and modified components.Source: Figure 18 · p. 63, Table 12 · p. 63
• Memory targets include episodic experience, semantic knowledge, and procedural strategies, as well as their representations, organization, and write, retrieval, update, and forgetting policies.Source: Appendix A · p. 64
• Workflow targets cover agent graphs and pipelines, task planning and decomposition, and routing, scheduling, and computational-resource allocation.Source: Appendix A · p. 64
• Harness and control targets include verification and reflection loops, multi-agent roles and protocols, and the scaffold code that implements orchestration and runtime coordination.Source: Appendix A · p. 64
• Tool-related targets range from available tools and their interfaces or implementations to reusable skills, macros, and rules for composing them.Source: Appendix A · p. 64
• Model targets include backbone weights, adapters and auxiliary modules, architecture, and persistent inference policies or configurations.Source: Appendix A · p. 64
• Training-system targets span objectives, optimizers, training schedules, data curricula, and search or meta-optimization procedures.Source: Appendix A · p. 65
• Other targets encompass evaluators and feedback, training data and environments, task-level artifacts, and full-system co-evolution or meta-RSI loops.These include judges, reward and fitness functions, verifiers, experience data, task generators, simulators, world models, solution code, algorithms, and scientific or design artifacts.Source: Appendix A · p. 65

• Table 13 organizes surveyed industrial systems by company archetype and improvement target.Source: Appendix B · p. 66

Company Archetypes in the Emerging RSI / Self-Improvement Landscape

• The table maps 72 companies and teams by archetype, with its RSI tag representing a B0–L5 scheme rather than each company’s own terminology.The snapshot is dated September 2026 and is organized by archetype, not geography.Source: Table 13 · p. 66, Company Archetypes in the Emerging RSI… · p. 66
• Agent Lightning, SkillOpt, coding agents, verifier pipelines, self-training evaluators, and HyperAgents illustrate improvement artifacts spanning harnesses, skills, code, evaluation, and meta-agents.The listed workflows include training from harness traces, editing skill files, generating and revising code, refining evaluation pipelines, evaluator self-training, and candidate L5 meta-agent modification.Source: Company Archetypes in the Emerging RSI… · p. 67
• Alibaba/Qwen, AgentWorld, and Qwen-Scope represent agent training through synthetic environments and data, while DeepSeek R1 and R-Zero illustrate reasoning bootstrapping and self-play.The fragments associate reasoning improvement with self-generated reasoning, verification, challenge generation, solving, voting, and policy training.Source: Company Archetypes in the Emerging RSI… · p. 67
• MiniMax M2, MaxHermes, MaxClaw, Kimi K3, Agent Swarm, and Hermes Agent target persistent or long-horizon agents through memory, skills, tool use, orchestration, and coordination.The entries describe persistent cloud agents and long-run execution alongside agent orchestration, tool-policy planning, and swarm coordination.Source: Company Archetypes in the Emerging RSI… · p. 67
• AutoGLM and AgentRL are listed as computer-use agents, with the accompanying fragments describing perception, planning, action, reinforcement learning, and virtual-phone training.The entries connect computer-use agents to environment training and policy weights.Source: Company Archetypes in the Emerging RSI… · p. 67
• Cogito v2 and IDA represent reasoning post-training, Poolside’s Model Factory automated model R&D, and Tinker and Inkling model customization.The fragments describe search, distillation, and iterative alignment for reasoning, and synthetic data, reinforcement learning, architecture evaluation, and code-execution RL for automated R&D.Source: Company Archetypes in the Emerging RSI… · p. 67

B. RSI-native / AI4AI companies

• Ricursive Intelligence targets AI-chip and compute-stack co-design, while Recursive Automated AI Research searches model-training recipes, kernels, and code through experiments and iteration.The table describes design search, verification, and iteration for Ricursive Intelligence, and branching and merging experiments for Recursive Automated AI Research.Source: B. RSI-native / AI4AI companies · p. 68
• Imbue Catalyst and Darwinian Evolver search populations of research recipes, code, and hypotheses, while another researcher entry describes code rewriting with evaluation and inheritance.The researcher entry labels this approach a candidate for level 5.Source: B. RSI-native / AI4AI companies · p. 68
• Sakana AI’s Darwin Godel Machine and AI Scientist are listed as self-editing agents, while Evolvent AI Org, RSIBench, and Terrarium center software-agent evolution.The latter entries emphasize evolving skills, memory, code, and environments through task feedback and refactoring.Source: B. RSI-native / AI4AI companies · p. 68
• MetaCircle’s ComfyResearch and OPHIS pursue AI4AI training workflows, while Frontis entries are labeled self-improving and EvoMap entries turn code assets into reusable genes or capsules.The EvoMap workflow is described as generating, testing, publishing, and reusing code assets.Source: B. RSI-native / AI4AI companies · p. 68

C. Autonomous R&D & scientific-discovery companies

• This subsection frames autonomous research and discovery as the primary product, with RSI relevance arising from automated experiment and knowledge-production loops.The catalog also includes autonomous AI-research data generation, research-agent evaluation, and scientific-model training entries.Source: B. RSI-native / AI4AI companies · p. 68, C. Autonomous R&D &… · p. 69
• Autonomous-science efforts span Periodic Labs’ hypothesis-and-experiment loop, FutureHouse’s bioinformatics evaluation, Edison Scientific’s autonomous science, and Lila Sciences’ discovery platform.The listed workflows include experiment design and lab runs, literature and experiment synthesis, scientific hypotheses and code, and lab execution.Source: C. Autonomous R&D &… · p. 69
• Axiom Math and Harmonic Aristotle target formal mathematics through conjecture, proof search, and verification, while Core Automation and Discovery Loop automate systems research and experimentation.Axiom Math’s listed representative works include Putnam 2025 and IMO 2026.Source: C. Autonomous R&D &… · p. 69
• Prime Intellect and Speedrun target model R&D with training recipes, optimizers, code, and GPU-run ablations, while Analemma and research-agent entries automate research workflows and experiments.The agent entries include evaluation harnesses, tool use, research traces, coding, and multi-agent experiments.Source: C. Autonomous R&D &… · p. 69

D. Agent optimization, evaluation & learning infrastructure

• Agent-improvement infrastructure includes skill revision, evaluator feedback, software test-and-repair, RL environments, and simulation and evaluation tooling.These entries describe infrastructure for feedback, skills, evaluation, simulators, training, and production-agent improvement.Source: D. Agent optimization, evaluation &… · pp. 69–70

E. Embodied, world-model & continual-adaptation companies

• Embodied and continual-adaptation efforts span VLA policy and action-data stacks, world models, test-time adaptation, and LoRA or skill-vector transfer; harness adaptation is marked a candidate.The entries also include real-robot deployment and tactile world-model work.Source: E. Embodied, world-model &… · pp. 70–71

F. Persistent-memory & personal-AI companies

• Persistent-memory and personal-AI systems center on cross-task memory and reusable skills, including long-term-memory retrieval and skill forging.The section characterizes these as substrates for adaptation across tasks.Source: F. Persistent-memory & personal-AI… · p. 71

Appendix C: RSI Across Applications

• Across applications, RSI workflows are compared by what they improve, how feedback validates updates, what components they change, and how those changes are retained.The review organizes evidence into science, embodied intelligence, software engineering, and healthcare, whose scenarios may overlap.Source: Appendix C · p. 72

C.1 S1: RSI for Science

• Scientific RSI uses evidence from computational or physical experiments to diagnose and improve the system that generates hypotheses and experiments, not just to find a better scientific result.The loop begins with a research challenge, proposes hypotheses and experiments, evaluates them, and uses evidence or falsification to diagnose the AI-for-science system.Source: §C.1 · p. 72
• Scientific RSI faces open-ended exploration, sparse and ambiguous feedback, and validity tied to assumptions and measurement, requiring provenance, scope, and uncertainty to travel with updates.The review groups evolving components as the scientific hypothesis module, experimental agent, and reflection system or improver.Source: §C.1 · p. 72

C.1.1 Evolving Scientific Hypothesis Modules

• HypoForge [158] distills critiques and execution outcomes into reusable skills, whereas EvoScientist [173] retrieves promising and failed research directions while keeping its models, memory schema, and update procedure fixed.HypoForge updates generation skills when direct empirical feedback is unavailable and refines testing skills using execution outcomes and attributed failures.Source: §C.1.1 · p. 72
• TTT-Discover [174] updates weights through test-time reinforcement learning, and the polymer framework retrains on simulation-evaluated candidates; both remain task- or domain-specific, supporting bounded rather than general cross-domain improvement.These systems use fixed training and selection procedures; current evidence does not show a general mechanism that progressively improves scientific hypothesis formulation across domains.Source: §C.1.1 · p. 72

C.1.2 Evolving Experimental Agents

• SciEvo’s Test-Time Tool Evolution [159] synthesizes, verifies, and reuses tools across 1,590 benchmark tasks, supported by 925 evolved tools.Source: §C.1.2 · p. 73
• CASCADE [176], S1-NexusAgent [177], and SkillFoundry [178] accumulate reusable scientific skills, making experimental agents’ action spaces persistent and extensible.SkillFoundry mines heterogeneous sources, packages executable skills, and validates, merges, or prunes them using execution and downstream utility; generation and admission rules remain externally specified.Source: §C.1.2 · p. 73
• Experience from 16 tasks improves performance on 17 held-out tasks in DrugSAGE [160], showing that inherited procedures and fixes transfer across tasks.DrugSAGE stores verified modeling procedures, training strategies, and recurring error fixes in cross-task memory.Source: §C.1.2 · p. 73
• STELLA [179], EarthLink [180], and OriGene refine research templates and workflows, while a fluid-control agent [182] revises both its definition and controller.Evaluation must distinguish improvements to the experimental agent from improvements to its scientific artifact, such as the controller.Source: §C.1.2 · p. 73
• Reliable experimental-agent evolution depends on validated reuse, with explicit skill preconditions, execution tests, provenance, versioning, and rollback guarding against out-of-context inheritance.Growth in a tool or memory library is weaker evidence than validated reuse on later tasks.Source: §C.1.2 · p. 73

C.1.3 Evolving Reflection Systems and Improvers

• An evolving improver updates its diagnostic, intervention, or promotion policy from earlier outcomes, unlike a fixed critic that merely revises the current output.Source: §C.1.3 · p. 73
• SIA [161] adapts agent components, CORAL [183] replaces fixed search heuristics with reflective asynchronous agents, and SAGA [184] learns executable objectives to steer search.SIA can modify scaffolds, tools, retry and search logic, or LoRA weights; CORAL externalizes discoveries as shared attempts, notes, and skills; SAGA adjusts concrete objectives under a fixed high-level goal.Source: §C.1.3 · p. 73

C.1.4 Toward Recursively Improving Science Agents

• Scientific RSI should improve hypothesis generation, experiment execution, and evidence interpretation so each cycle yields findings and reusable improvements to future discovery.Current AI4Sci systems span widespread B0 output refinement, demonstrated L1 persistent updates, and an L2 frontier; aspects of L3 exist, but end-to-end L4 adaptation and L5 evolution of the improver remain largely unexplored.Source: §C.1.4 · pp. 73–74
• Progress toward scientific RSI requires reliable outcome attribution, selective cross-task transfer, and evaluation of future research gains, while preserving provenance, reproducibility, and scientific and safety constraints.The challenge applies as increasingly powerful parts of the improvement process become adaptive.Source: §C.1.4 · p. 74

C.2 S2: RSI for Embodied Intelligence

• Embodied RSI must address endogenous experience, cross-component credit assignment, and safe inheritance because physical trials are difficult to reproduce, reverse, or reset.Its loop organizes tasks and capabilities, executes actions, interprets outcomes, and writes validated feedback back into components for subsequent interaction.Source: §C.2 · p. 74

C.2.1 Evolving Environments and Curricula

• POET [65], EnvGen [185], GenEnv [186], OMNI-EPIC [187], and SimWorld Studio [188] use learner progress to evolve curricula, task distributions, or generators of future embodied experience.Approaches range from co-evolving environment–agent pairs and LLM-designed curricula to generated executable worlds and progressively harder Gym-compatible environments.Source: §C.2.1 · p. 74

C.2.2 Evolving Skills, Memory, and Agent Harnesses

• Voyager [84], LRLL [189], EmbodiSkill [190], SHAPER [111], ASPIRE [191], and ENPIRE [192] extend improvement from skill libraries and evidence-based repair to evolving harnesses, transferable fixes, and real-world policy or infrastructure changes.This system-level evolution changes not only what the agent knows, but also how capabilities are selected, composed, and improved.Source: §C.2.2 · p. 75

C.2.3 Evolving Policies and Action Models

• Self-Improving Embodied Foundation Models [193], MEDAL++ [194], Q-Planning [195], and SERP [196] learn from interaction, but their update algorithms, rewards, and permitted model components remain specified in advance.Methods support practice with limited supervision, learning task completion and undoing, training value functions from deployment trajectories, or adapting action models after navigation failures.Source: §C.2.3 · p. 75

C.2.4 Evolving World Models and Evaluators

• VLAW [197], World-VLA-Loop [162], Motus2 [198], and SAIL [199] improve world models, evaluators, or policies from interaction, closing a bounded loop whose objectives and updates remain externally fixed.Their feedback can update the policy, harness, or curriculum for the next round, but the systems remain bounded model–policy co-evolution rather than unrestricted recursive improvement.Source: §C.2.4 · p. 75

C.2.5 Toward Recursively Improving Embodied Agents

• Embodied RSI should generate frontier-level tasks, assemble relevant memories, skills, tools, and models, interpret outcomes, and pass verified improvements into the policy, world model, or harness.This loop is intended to improve both current behavior and future learning across interactions.Source: §C.2.5 · pp. 75–76
• Embodied systems show partial progress: B0 reflection is widespread; L1 policy/model updates [193] [194] [199], L2 component updates [111] [190] [191], L3 state-guided experience [84] [189], and L4 co-evolution [65] [185] [186] [188] are demonstrated.ENPIRE [192] approaches bounded L5 by revising policies, training procedures, and infrastructure, but its interfaces, evaluators, and safety constraints remain externally specified.Source: §C.2.5 · p. 76
• True embodied RSI requires informative experience acquisition, sensorimotor outcome attribution, physical-condition validation, cross-task and cross-embodiment robustness, provenance, uncertainty, rollback, and safe autonomous improvement of its harness and evaluation.Current methods demonstrate these abilities separately but have not integrated them into a safe, persistent, autonomous real-world loop.Source: §C.2.5 · p. 76

C.3 S3: RSI for Software Engineering

• Software engineering offers a concrete RSI setting because both products and developing agents are executable code; agents use repository feedback to change persistent skills, workflows, or implementations, then reuse accepted changes.Feedback can include test failures, runtime traces, or code review.Source: §C.3 · p. 76
• Software RSI must generate valuable, verifiable tasks targeting capability gaps and assess whether changes support further improvement, since immediate benchmark gains may stagnate later search while weaker variants may yield stronger descendants.The challenge is to optimize both present effectiveness and future improvement capacity.Source: §C.3 · p. 76
• Coordinating the evolution of agent implementations and harnesses, software-engineering experience and collaboration, and the improvement process itself while preserving the outer evaluation and objective contract remains a challenge.Source: §C.3 · p. 76

C.3.1 Evolving Coding-Agent Implementations and Harnesses

• SICA [200] edits its own Python codebase using prior versions and benchmark results, but its benchmark, utility function, and sandbox remain externally fixed.The editable agent can improve its coding and later self-modification through changes to tools, subagents, or context management.Source: §C.3.1 · p. 76
• Self-Harness [23], Agentic Harness Engineering [63], and Ouroboros [18] evolve harnesses through tested or reviewed changes that persist into later executions.Their methods use execution traces or trajectory evidence, regression testing or predicted-effect checks, and deployment of accepted changes to tools, context, prompts, or implementation.Source: §C.3.1 · pp. 76–77

C.3.2 Evolving Software-Engineering Experience, Skills, and Collaboration

• SWE-Exp [60] retrieves lessons from successful and failed trajectories, while CODESKILL [201] updates a compact bank of procedural skills for future agents.Both retain reusable development knowledge across tasks, but their extraction and update procedures do not themselves improve through use.Source: §C.3.2 · p. 77
• EvoMAC [202] revises multi-agent role prompts and communication links using test feedback and textual back-propagation, with reported updates occurring during each task’s test time.Source: §C.3.2 · p. 77

C.3.3 Evolving the Improvement Process

• DGM searches multiple coding-agent lineages, while HGM allocates evaluation by estimated descendant improvement potential rather than current benchmark performance alone.DGM archives variants and evaluates descendants, allowing weaker variants to serve as stepping stones; HGM estimates metaproductivity from descendant lineages.Source: §C.3.3 · p. 77
• Group-Evolving Agents [203] shares modifications and failures across parent agents, DarwinX [204] filters harnesses for broader nonregressing task coverage, and HELIX [205] alternates harness and model updates.HELIX uses harness evolution to generate verified trajectories for model updates, then rebuilds the harness for the changed model.Source: §C.3.3 · p. 77

C.3.4 Toward Recursively Improving Software-Engineering Agents

• Software-engineering RSI follows a loop of development challenge, agent modification, repository-grounded trial, regression-aware selection, and versioned inheritance.Repository feedback triggers persistent updates evaluated on current and held-out tasks before traceable, reversible versions are inherited.Source: §C.3.4 · p. 77
• Software agents show partial RSI: L1 cross-task learning and L2 bounded change are established, L3 task generation is emerging, L4 is largely absent, and full L5 remains undemonstrated.L1 includes cross-task experience and skill learning [60] [201]; L2 selects harness, tool, or workflow changes under fixed objectives and evaluators [23] [63] [204]. L3 learner-conditioned generation appears in Self-play SWE-RL and Socratic-SWE [207]; SICA and DGM show bounded L5 characteristics [200], while HGM adds lineage-based meta-selection without changing its selection rule. True RSI also requires robust specifications, attributable improvements, reliable transfer, and externally protected user intent, safety, auditability, and rollback.Source: §C.3.4 · pp. 77–78

C.4 S4: RSI for Healthcare

• Healthcare RSI turns clinical feedback into persistent updates to knowledge, strategies, or workflows, but risky trial and error, delayed or confounded outcomes, and setting-specific effects require oversight and governed inheritance.Updates validated in one population or institution cannot safely transfer without recording provenance, scope, and uncertainty.Source: §C.4 · p. 78

C.4.1 Evolving Clinical Memory and Knowledge

• MedAgent-Zero [163], Agent Mental Clinic [208], and MedAgentSim persist case experience, while DxEvolve [210], GSEM [211], and Evo-MedAgent [212] refine memory structure, applicability, or tool trust.Most evidence uses simulated or retrospective cases with externally supplied labels; validation of memory evolution in real-world cases remains lacking.Source: §C.4.1 · p. 78

C.4.2 Evolving Clinical Reasoning Strategies

• EvoClinician [164] evolves diagnostic policy by grading each question or examination for clinical yield and resource cost, then using feedback to revise prompts and memory before later cases.The inherited policy can determine which symptom to clarify or examination to prioritize.Source: §C.4.2 · p. 78
• EvoMDT updates specialist prompts, consensus weights, and retrieval scope, while PsychAgent [214] adds practice-grounded counseling skills and internalizes selected skills through rejection fine-tuning.These approaches evolve multi-agent coordination or an agent’s skill repertoire using expert assessments, outcomes, or historical counseling trajectories.Source: §C.4.2 · p. 79

C.4.3 Evolving Healthcare Tools and Workflows

• MACRO [165] turns verified imaging patterns into composite tools, while SkeMex [215] distills procedural skills and manages them according to context-dependent utility.Their evolving action components are promoted, merged, updated, or removed based on demonstrated utility in later tasks.Source: §C.4.3 · p. 79
• TissueLab [216] uses expert corrections to refine imaging workflows, while HealthFlow retains safeguards and reusable EHR procedures; inherited workflows still require validation and reversibility.HealthFlow evaluates execution or methodological defects before a reflector validates, updates, or retires stored experience.Source: §C.4.3 · p. 79

C.4.4 Toward Recursively Improving Healthcare Agents

• Healthcare RSI should improve task performance while preserving clinical validity and safety, with each episode producing a validated update to memory, reasoning, or action repertoire for safe reuse.The resulting capability should be further evaluated in subsequent cases.Source: §C.4.4 · p. 79
• B0 reflection is common but nonpersistent; L1 inherits reusable clinical knowledge [208] [210], L2 is the strongest current frontier [164] [165] [211] [212] [215], while L3/L4 remain bounded and L5 is largely unexplored [163] [216].L3 elements shape evidence acquisition but leave patient populations and task distributions externally specified; EvoPatient explores L4-style co-evolution, while end-to-end L4 adaptation from real outcomes remains largely unexplored.Source: §C.4.4 · p. 79
• Progress requires reliable attribution of longitudinal outcomes, safe improvement feedback without unrestricted clinical exploration, selective transfer across populations and institutions, and governed validation, monitoring, traceability, and reversibility of updates.Persistent updates should be prospectively monitored and reversible when their assumptions or safety guarantees no longer hold.Source: §C.4.4 · p. 79