Pith

open record

sign in

arxiv: 2607.07663 · v1 · pith:QNTXHFKR · submitted 2026-07-08 · cs.AI

Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops

Mingguang Chen , Licheng Wang , Bo Qu This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 reserved 2026-07-09 03:07 UTCglm-5.2pith:QNTXHFKRrecord.jsonopen to challenge →

Figure 1
Figure 1. Figure 1: the two-axis taxonomy, with representative systems per cell. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] reproduced from arXiv: 2607.07663
classification cs.AI
keywords recursive self-improvementverification hierarchyself-evaluationAI safetymodel collapseLLM self-trainingevaluator co-evolutionAI governance
0
0 comments X

The pith

Self-improving AI lives or dies by its verifier

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This survey of 1,250 papers argues that the entire landscape of AI self-improvement — from inference-time output revision to autonomous research agents — is governed by a single organizing principle: the strength of any self-improvement loop is bounded by the reliability of the signal that tells it whether it got better. The authors propose a two-axis taxonomy separating what a system improves (its outputs, its weights, its evaluator, or the research process itself) from how closed the loop is (human-in-the-loop to fully autonomous). Across all four categories, they observe that demonstrated improvement tracks a verification hierarchy: formal proof checkers at the top, execution feedback next, learned judges below that, and intrinsic self-assessment at the bottom. The characteristic failure modes — self-confirming loops, model collapse, diversity collapse — arise precisely when a system's evaluator sits too low on this hierarchy relative to its ambition. The survey's most consequential claim is that the bottleneck preventing fully autonomous AI research (choosing which problems matter) and the bottleneck preventing reliable self-evaluation are the same bottleneck, and that this bottleneck sits at the top of the verification hierarchy where no current system can operate without human judgment.

Core claim

The central discovery is the verification hierarchy as an empirical regularity across the self-improvement literature. Formal verifiers (proof checkers, type systems) are sound by construction and permit indefinite iteration without accepting false improvements. Execution feedback (tests, benchmarks) is reliable but incomplete. Learned judges (reward models, LLM-as-judge) are bounded by the judge's own competence and are themselves gameable. Intrinsic signals (confidence, self-consistency) are cheapest and most gameable. The authors observe that every demonstrated self-improvement success in the corpus sits at the top two rungs (code, math, formal verification), while every persistent gap —,

What carries the argument

The verification hierarchy: a four-level ordering of evaluator reliability from formal verifiers (strongest) through execution feedback, learned judges, to intrinsic self-assessment (weakest), which the authors claim predicts both where self-improvement works and where it fails.

If this is right

  • If the verification hierarchy is real, the path to more autonomous AI research runs through evaluator engineering, not raw model capability — the binding constraint is building systems that can reliably judge open-ended quality, not systems that are merely smarter.
  • Governance of self-improving AI becomes a measurement problem: regulators would need to audit not what a system produces but what evaluator it runs against and where that evaluator sits on the hierarchy.
  • The distinction between bounded self-refinement (convergent, evaluable, already industrial) and open-ended RSI (divergent, unverified, still theoretical) gives policymakers a concrete vocabulary for separating near-term engineering risks from speculative takeoff scenarios.
  • Evaluator co-evolution — the emerging practice of having systems improve their own verifiers alongside their policies — is identified as the pivotal empirical question: it either escapes the self-confirming loop or gives it a second story, and the answer determines whether closed-loop self-improvement stabilizes or compounds bias.
  • The field's smallest category (foundations, limits, and safety at 60 of 1,250 papers) represents the largest mismatch between stated stakes and research investment, suggesting the academic incentive structure is underweighting exactly the questions that governance depends on.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the verification hierarchy generalizes beyond the corpus, one would predict that self-improvement in domains with cheap verifiers (software, formal math, competitive programming) will continue to advance rapidly while domains without them (scientific research, creative work, strategic planning) will stall regardless of model scale — a domain-specific ceiling on RSI.
  • The exchange rate of grounding — the minimum fraction of external signal needed to prevent collapse — is a measurable quantity that could be established experimentally by systematically varying the ratio of exogenous to self-generated data in training loops and observing the collapse threshold, turning a theoretical question into an empirical one.
  • If research taste (identifying which problems are worth working on) is the top rung of the hierarchy and is beginning to acquire a formal theory, then benchmarks for interestingness or problem-selection quality would be the single highest-leverage intervention for the field, as they would convert the last human-in-the-loop bottleneck into a measurable target.
  • The A-Evolve-Training episode, where a system detected its own proxy metric corrupting and revised its search policy, may represent the first field observation of a system climbing the verification hierarchy autonomously — if this capability generalizes, it would be evidence that evaluator self-correction is achievable, weakening the claim that the verification bottleneck is permanent.

Load-bearing premise

The verification hierarchy is described as a qualitative pattern observed across the corpus, not a measured law. Because the corpus was assembled by seed queries that may systematically over-represent verifiable domains (code, math) where self-improvement works and under-represent domains where weaker signals succeed, the hierarchy could partly reflect sampling bias rather than a genuine empirical regularity about AI self-improvement.

What would settle it

Find a substantial body of cases where self-improvement loops succeed reliably using only intrinsic signals (model confidence, self-consistency) without external verification — this would break the claimed hierarchy by showing the bottom rung can sustain improvement at scale.

Figures

Figures reproduced from arXiv: 2607.07663 by Bo Qu, Licheng Wang, Mingguang Chen.

Figure 2
Figure 2. Figure 2: semantic map of the 1,250-paper corpus (TF-IDF abstracts, SVD + t-SNE projection; [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: deployment-time self-evolution ordered by persistence — refined outputs evaporate with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: the training-time self-iteration loop. The five paradigms of §4 differ mainly in who [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: the verification hierarchy. Signal reliability rises toward the top while task coverage [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Seed-corpus quarterly growth through 2026Q2, showing log-scaled paper counts and [PITH_FULL_IMAGE:figures/full_fig_p026_6.png] view at source ↗
read the original abstract

AI systems increasingly participate in their own improvement: revising their outputs, adapting their own harnesses during deployment, training on data they generate, and, increasingly, conducting AI research itself. This literature is described under a vocabulary ("self-refine," "self-reward," "self-play," "self-evolve") that conflates fundamentally different ambitions. We survey 1,250 arXiv papers (2024-2026) along two axes: what the system improves -- its behavior in deployment, its policy through training, its evaluator, or the research process itself -- and the degree of loop closure (human-in-the-loop to fully closed). The taxonomy separates bounded self-refinement -- convergent, evaluable, and already industrial practice -- from open-ended recursive self-improvement (RSI), which remains bounded by grounding requirements, collapse dynamics, and compute constraints on every measured axis. Its distinctive feature is a dedicated category for self-evaluation: every improvement loop is a claim that some signal can substitute for human judgment. We survey the evaluator design space -- judges, process reward models, verifiers, rubrics, meta-evaluation -- order the signals into a verification hierarchy from formal verifiers (strongest) to intrinsic self-assessment (weakest), and observe that demonstrated self-improvement strength tracks this hierarchy, that its failure modes (self-confirming loops, model collapse, diversity collapse) follow from its violations, and that the "research direction-setting" bottleneck keeping humans in the loop sits at the top of that hierarchy. We connect the technical literature to the theory of RSI limits and to the safety and governance questions raised by frontier-lab accounts of closing the loop, and identify governance-grade measurement of self-improvement as the field's most underpopulated niche.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 8 minor

Summary. This manuscript surveys 1,250 arXiv papers (2024–2026) on AI self-improvement, organizing them along two axes: what the system improves (deployment-time behavior, training-time policy, evaluator, or research process) and the degree of loop closure (human-in-the-loop to fully closed). The central organizing claim is a 'verification hierarchy'—that demonstrated self-improvement strength tracks the reliability of the evaluation signal, from formal verifiers (strongest) to intrinsic self-assessment (weakest)—and that characteristic failure modes (self-confirming loops, model collapse, diversity collapse) follow from violations of this hierarchy. The survey covers four technical categories (deployment-time self-evolution, training-time self-iteration, self-evaluation, auto research) plus a foundations/limits/safety family, and connects the technical literature to RSI theory and governance questions. The corpus, classification scripts, and per-paper assignments are released as reproducible artifacts.

Significance. The paper addresses a genuine gap: no existing survey spans the full spectrum from bounded self-refinement to open-ended RSI while treating self-evaluation as a load-bearing pillar. The two-axis taxonomy is a useful organizing contribution, and the verification hierarchy, if valid, reframes the RSI debate from speculation about capability growth into a measurement program focused on evaluator reliability. The authors ship reproducible corpus construction scripts and per-paper assignments—a concrete strength. The argumentative skeleton rests on verified anchor works (STaR, Self-Refine, FunSearch, Huang et al.'s negative result on intrinsic self-correction) and the diagnostic literature is treated as load-bearing rather than incidental. The connection to Anthropic's autonomy continuum and the identification of governance-grade measurement as an underpopulated niche are well-motivated.

major comments (3)
  1. §5.2 and §5.4: The verification hierarchy is the paper's central load-bearing empirical claim, but it is validated only qualitatively within a corpus whose construction was partly shaped by the taxonomy that encodes the hierarchy. The supplemental harvest (§2.3) targeted 379 papers in 'directions the taxonomy makes first-class,' including self-evaluation methods. The paper's organizing premise—'every improvement loop is a claim that some signal can substitute for human judgment' (§5)—predisposes the analysis to find that evaluator quality is the key variable. If every loop is framed as depending on an evaluator, and loops with stronger evaluators are observed to work better, the conclusion is partly built into the lens. The authors explicitly caveat this ('a qualitative pattern we observe throughout the corpus, not a measured law'), but the caveat does not resolve the structural concern.
  2. §5.4: The claim that 'the direction-setting bottleneck and the verification bottleneck are the same bottleneck' is asserted rather than demonstrated. The Anthropic essay [5] identifies research direction-setting as a distinct bottleneck (choosing which problems matter), while the verification hierarchy concerns signal reliability for already-specified tasks. The conflation of these two different bottlenecks—problem selection and solution verification—is load-bearing for the paper's framing but is not separately argued. The within-domain evidence (Huang et al. [27], Mirror Loop [140], Lin [92]) supports the verification claim but does not address whether direction-setting reduces to verification.
  3. §2.3: The corpus construction has a recency bias (74% from 2026) that interacts with the verification hierarchy claim in a way that is not fully addressed. If the field has recently concentrated on verifiable domains (code, math) where self-improvement works, the hierarchy could partly reflect where current research activity happens to be dense rather than a stable empirical regularity. The authors acknowledge this for citation counts but not for the hierarchy claim itself. A simple robustness check—reporting the hierarchy's support stratified by publication year—would help distinguish a real regularity from a recency artifact.
minor comments (8)
  1. Figure 1 is referenced as laying out the 4×3 grid with representative systems, but the figure content is not visible in the manuscript text. Ensure the figure is properly rendered and legible in the final version.
  2. Figure 2 (semantic map) uses TF-IDF abstracts with SVD + t-SNE projection. The axes are described as 'arbitrary embedding dimensions' but no clustering quality metric (e.g., silhouette score) is reported to support the claim that 'Auto Research and foundations form coherent regions.'
  3. Table 1 reports '82% posted in 2026' for the self-evaluation category, which is notably higher than other categories (57–76%). Given that the supplemental harvest specifically targeted self-evaluation methods, this percentage may reflect sampling design rather than field dynamics. A footnote clarifying this would help.
  4. §4.3: The claim that on-policy self-distillation 'did not exist before 2026' is strong; the authors should verify whether precursor work (e.g., earlier self-distillation variants) existed under different names.
  5. §7.1: The Whitfill and Wu result [189] is described as 'the empirical crux of the RSI-feasibility debate,' but the two specifications diverging (substitutes vs. complements) means the result is inconclusive rather than a crux. The framing should be softened.
  6. The reference list includes several papers from 2026 with arXiv IDs starting '2607' (July 2026), which is the same month as this manuscript's submission. The authors should verify these are not concurrent submissions that could create citation circularity.
  7. §3.5: The term 'harness' is defined in §2.1 but used extensively before its definition appears in the reading order. Consider forward-referencing the definition in §3.
  8. The abstract states '1,250 arXiv papers' while §2.3 describes 871 seed + 379 supplemental = 1,250. This is consistent but the abstract could note the two-stage construction for clarity.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for a careful and constructive report. The referee identifies three major concerns: (1) a potential circularity in the verification hierarchy claim arising from the taxonomy-guided corpus construction, (2) an insufficiently argued conflation of the direction-setting bottleneck with the verification bottleneck, and (3) an unaddressed interaction between corpus recency bias and the hierarchy claim. We agree that all three points identify genuine weaknesses in the current manuscript and will revise accordingly. Specifically, we will (a) add an explicit methodological caveat about the taxonomy-corpus interaction and reframe the hierarchy's epistemic status, (b) separate the two bottlenecks and argue the reduction rather than asserting it, and (c) add a stratified robustness check by publication year. We cannot fully resolve the structural circularity concern—it is inherent to any taxonomy-guided survey—but we can make it fully transparent and distinguish what the hierarchy claims to be (an organizing empirical pattern) from what it does not claim to be (an independently validated law).

read point-by-point responses

  1. Referee: §5.2 and §5.4: The verification hierarchy is the paper's central load-bearing empirical claim, but it is validated only qualitatively within a corpus whose construction was partly shaped by the taxonomy that encodes the hierarchy. The supplemental harvest (§2.3) targeted 379 papers in 'directions the taxonomy makes first-class,' including self-evaluation methods. The paper's organizing premise—'every improvement loop is a claim that some signal can substitute for human judgment' (§5)—predisposes the analysis to find that evaluator quality is the key variable. If every loop is framed as depending on an evaluator, and loops with stronger evaluators are observed to work better, the conclusion is partly built into the lens.

    Authors: The referee is correct that there is a structural circularity here, and we had not stated it with sufficient force. The concern has two layers, and we address each. First, the corpus-construction layer: the supplemental harvest (§2.3) targeted self-evaluation methods, test-time training, and zero-data self-play because the taxonomy made them first-class. This means the corpus is not an independent sample with respect to the taxonomy—it was partly shaped by it. We will add an explicit methodological caveat in §2.3 stating this directly and noting that the corpus proportions cannot be read as unbiased estimates of the field's composition. Second, the conceptual layer: the premise that 'every improvement loop is a claim that some signal can substitute for human judgment' does frame every loop as evaluator-dependent, which makes the finding that evaluator quality matters partly built into the lens. We acknowledge this honestly. However, we note that the premise is not merely an assumption—it is a logical observation: any improvement loop requires a criterion for 'better,' and that criterion is an evaluator by definition. What is not built into the lens is the empirical claim that stronger evaluators (formal verifiers) produce more durable improvement than weaker ones (intrinsic self-assessment). That claim is supported by independent anchor evidence (Huang et al. [27], the Mirror Loop [140], Lin [92], FunSearch [3], AlphaEvolve [4]) that was not selected by the taxonomy but by its diagnostic quality. Still, the referee is right that the qualitative pattern across the full corpus is observed through the taxonomy's lens, and we will reframe the hierarchy's epistemic status accordingly. We will (a) add a paragraph in §5.2 explicitly acknowledging the taxonomy-corpus circularly revision: partial

  2. Referee: §5.4: The claim that 'the direction-setting bottleneck and the verification bottleneck are the same bottleneck' is asserted rather than demonstrated. The Anthropic essay [5] identifies research direction-setting as a distinct bottleneck (choosing which problems matter), while the verification hierarchy concerns signal reliability for already-specified tasks. The conflation of these two different bottlenecks—problem selection and solution verification—is load-bearing for the paper's framing but is not separately argued. The within-domain evidence (Huang et al. [27], Mirror Loop [140], Lin [92]) supports the verification claim but does not address whether direction-setting reduces to verification.

    Authors: The referee is correct. The claim as stated conflates two distinct bottlenecks: (1) verifying that a solution to a specified problem is correct, and (2) selecting which problems are worth solving. The Anthropic essay [5] treats these as distinct, and our evidence base (Huang et al. [27], Mirror Loop [140], Lin [92]) supports only the first claim. We will revise §5.4 to separate the two bottlenecks and argue the relationship rather than asserting identity. The argument we can honestly make is narrower: direction-setting is a harder problem than verification, and it sits at or above the top of the verification hierarchy because it requires evaluating not just whether a solution is correct but whether a problem is worth posing—a judgment for which no formal verifier exists. This makes direction-setting a superset of the verification problem, not the same problem. The evidence for this narrower claim comes from §6.3's diagnostic literature (ScienceAgentBench [177], ResearchArena [178], the integrity benchmarks [143]) and from Herrmann and Schmidhuber's formalization of interestingness as a bottleneck [190], which identifies problem selection as requiring inductive heuristics that go beyond verification. We will rewrite the claim to state that direction-setting subsumes the verification bottleneck and adds the further difficulty of evaluating problem importance, for which no trustworthy automated signal currently exists. We will remove the assertion of identity and replace it with this argued relationship. revision: yes

  3. Referee: §2.3: The corpus construction has a recency bias (74% from 2026) that interacts with the verification hierarchy claim in a way that is not fully addressed. If the field has recently concentrated on verifiable domains (code, math) where self-improvement works, the hierarchy could partly reflect where current research activity happens to be dense rather than a stable empirical regularity. The authors acknowledge this for citation counts but not for the hierarchy claim itself. A simple robustness check—reporting the hierarchy's support stratified by publication year—would help distinguish a real regularity from a recency artifact.

    Authors: This is a fair and actionable suggestion. We will add the requested stratified robustness check: for each rung of the verification hierarchy, we will report the distribution of supporting evidence across publication years (2024, 2025, 2026). This will let readers assess whether the hierarchy is driven by the 2026 mass or holds across the corpus's temporal span. We expect the hierarchy to be visible in the 2024 anchor works (STaR [7], Self-Refine [6], Huang et al. [27], FunSearch [3]) and in the 2025 diagnostic literature (Mirror Loop [140], Lin [92]), not only in the 2026 wave, but the referee is right that this should be shown rather than assumed. We will also add an explicit caveat in §5.2 noting that the hierarchy's support is concentrated in verifiable domains (code, math, formal methods) and that its applicability to non-verifiable domains is inferred by absence (the lack of demonstrated durable self-improvement without external signal) rather than by positive evidence of failure. The recency interaction is real: if the field's attention has shifted toward verifiable domains, the hierarchy could partly reflect where activity is dense. The stratified check will help distinguish these possibilities, though we acknowledge it cannot fully separate a genuine regularity from a field-wide attention shift—both could produce the same stratified pattern. We will state this limitation explicitly. revision: yes

standing simulated objections not resolved
  • The structural circularity between taxonomy-guided corpus construction and the verification hierarchy claim cannot be fully resolved. The taxonomy shapes the corpus, and the corpus is used to validate a claim encoded in the taxonomy. We can make this fully transparent, reframe the hierarchy's epistemic status as a qualitative pattern rather than a measured law, and point to independent anchor evidence, but we cannot make the corpus independent of the taxonomy without rebuilding it from scratch using taxonomy-free queries—which would lose the targeted coverage of self-evaluation that makes the survey distinctive.

Circularity Check

2 steps flagged

Mild self-definitional framing (improvement defined via evaluator → evaluator quality is the bottleneck), but transparently acknowledged and externally grounded; no self-citation chain.

specific steps
  1. self definitional [§2.1 (definition of Self-improvement) and §5 (verification hierarchy claim)]
    "Self-improvement. A system participates in producing a better version of itself or of its own outputs, where “better” is defined by some evaluator. The definitional dependence on an evaluator is not pedantry; it is the source of every failure mode in §5."

    The paper defines self-improvement as inherently dependent on an evaluator ('better is defined by some evaluator'), then derives that evaluator quality is the universal bottleneck ('the loop's ceiling is exactly the quality of that substitute'). If 'improvement' is defined as 'improvement against an evaluator,' then 'evaluator quality bounds improvement quality' follows partly by definition. However, the paper is explicitly transparent about this ('not pedantry; it is the source of every failure mode'), uses it as a deliberate framing device rather than a hidden derivation, and the specific empirical content—the ordering of the hierarchy (formal verifiers > execution > learned judges > intrinsic signals) and the three failure modes—is supported by independent external work (Huang et al. [2

  2. fitted input called prediction [§2.3 (corpus construction) and §5.2 (hierarchy observation)]
    "a targeted supplemental harvest of 379 papers in directions the taxonomy makes first-class: self-evaluation methods (judges, process reward models, verifiers, rubrics, meta-evaluation), test-time training, and zero-data self-play. [...] The empirical regularity across all four categories—a qualitative pattern we observe throughout the corpus, not a measured law—is that demonstrated self-improvement strength tracks this hierarchy."

    The taxonomy elevates self-evaluation to a first-class category, the supplemental harvest fills that category with 318 papers, and then the verification hierarchy is observed across all categories including the one constructed to embody it. This creates a mild framing-observation loop: the lens (taxonomy) partly shapes the data (supplemental harvest), and the conclusion (hierarchy) is read from data shaped by the lens. However, the paper is transparent about this ('the raw counts partly mirror our own sampling'), explicitly labels the hierarchy as 'not a measured law,' and the within-domain evidence supporting the hierarchy (Huang et al. [27], Lin [92], Mirror Loop [140], Shumailov et al. [144]) is drawn from independent external work, not from the corpus construction itself. The circulari

full rationale

This is a survey paper, not a formal derivation, so the circularity patterns targeting fitted-parameter predictions or self-citation chains largely do not apply. No self-citations exist in the reference list (authors Chen, Wang, Qu cite none of their own prior work), ruling out patterns 3–5. The two mild concerns are: (1) a self-definitional framing where 'improvement' is defined via an evaluator and then evaluator quality is found to be the bottleneck—transparently acknowledged by the paper as a deliberate framing choice; and (2) a corpus-construction loop where the taxonomy shapes the supplemental harvest which then supports the hierarchy observed across categories—transparently acknowledged as 'not a measured law' with corpus proportions explicitly noted to 'partly mirror our own sampling.' In both cases the paper is candid about the limitation, and the specific empirical claims (hierarchy ordering, failure modes) are grounded in independent external citations. The circularity is real but minor and self-disclosed, warranting a score of 2.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 2 invented entities

The survey introduces no new physical entities or fitted constants. The free parameters are methodological choices (query caps, harvest targets, classification rules) that shape the corpus. The axioms are domain assumptions about the nature of self-improvement and evaluator reliability, all stated explicitly and reasonable within the field. The invented entities are organizational constructs (the hierarchy, the taxonomy) rather than postulated physical objects, and both have falsifiable handles.

free parameters (3)
  • seed query depth caps = not specified
    Per-query caps on arXiv harvest affect corpus composition; the authors acknowledge these favor recent, high-volume threads.
  • supplemental harvest target directions = self-evaluation, test-time training, zero-data self-play
    The three directions chosen for supplemental harvest were determined by the taxonomy, creating a feedback between the framework and the corpus.
  • keyword classification rules = not fully specified in text
    Rule-based classification moved 89 papers off thread defaults; the rules themselves are in released scripts but their design is a free parameter of the method.
axioms (4)
  • domain assumption Every self-improvement loop depends on an evaluator that defines 'better.'
    Stated in §2.1 ('Self-improvement' definition) and §5 opening. This is the foundational premise of the survey's argument. It is a reasonable domain assumption but is not independently proven; it is asserted as definitional.
  • domain assumption The verification hierarchy (formal verifiers > execution feedback > learned judges > intrinsic signals) is a meaningful ordering of evaluator reliability.
    Introduced in §5.2. The ordering is intuitive and supported by cited examples, but the paper explicitly calls the tracking claim 'a qualitative pattern we observe throughout the corpus, not a measured law.' The hierarchy itself is treated as an axiom for organizing the survey.
  • domain assumption Bounded self-refinement and open-ended RSI are categorically distinct phenomena.
    Stated in §2.1 ('Bounded self-refinement vs. open-ended RSI'). This distinction is the survey's 'central cut' and is assumed rather than derived. The theory literature in §7 is cited as support, but the categorical distinction is an organizing premise.
  • domain assumption The arXiv corpus is a representative sample of the self-improvement literature.
    Implicit in the survey methodology (§2.3). The authors acknowledge the corpus is 'a sample, not a census' and note publication-censoring effects for industrial RSI practice, but the survey's claims depend on the sample being informative about the field's structure.
invented entities (2)
  • Verification hierarchy (4-level) independent evidence
    purpose: Orders evaluator signals from formal verifiers (strongest) to intrinsic self-assessment (weakest) to explain why some self-improvement loops work and others fail.
    Not a new physical entity but a new organizational construct. It is supported by cited empirical evidence (Huang et al. [27], Lightman et al. [130], Shumailov et al. [144]) and makes a falsifiable prediction: systems using higher-rung verifiers will show stronger, more stable self-improvement. The prediction is qualitative and not yet quantitatively tested in the paper.
  • Two-axis taxonomy (4 categories × 3 closure levels) independent evidence
    purpose: Classifies self-improvement methods by what they improve and who validates the improvement.
    An organizational framework, not a physical entity. It is validated by the corpus mapping (1,250 papers classified) and makes the field's structure visible. The taxonomy absorbed two paradigm arrivals (OPSD, zero-data self-play) without modification, which the authors cite as evidence of durability.

pith-pipeline@v1.1.0-glm · 38624 in / 3718 out tokens · 400979 ms · 2026-07-09T03:07:35.745829+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

205 extracted references · 205 canonical work pages · 159 internal anchors

  1. [1]

    Speculations concerning the first ultraintelligent machine

    Irving John Good. Speculations concerning the first ultraintelligent machine. Advances in Computers, 6:31–88, 1966. doi: 10.1016/S0065-2458(08)60418-0

    work page doi:10.1016/s0065-2458(08)60418-0 1966

  2. [2]

    Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements

    Jürgen Schmidhuber. Gödel machines: Self-referential universal problem solvers making provably optimal self-improvements. arXiv preprint cs/0309048 , 2003. URL https://arxiv.or g/abs/cs/0309048

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  3. [3]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models. Nature, 625:468–475, 2024. doi: 10.1038/s41586- 023-06924-6. ...

    work page doi:10.1038/s41586- 2024

  4. [4]

    Alexander Novikov, Ngân V u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and algor...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    Recursive self-improvement

    Anthropic. Recursive self-improvement. Anthropic Institute blog post, 2026. URL https: //www.anthropic.com/institute/recursive-self-improvement. Published May 2026

    work page 2026

  6. [6]

    Self-Refine: Iterative Refinement with Self-Feedback

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegr- effe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bod- hisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback. In Advances in Neural Information Processin...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstrapping reasoning with reasoning. In Advances in Neural Information Processing Systems (NeurIPS), 2022. URL https://arxiv.org/abs/2203.14465

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  8. [8]

    Self-Rewarding Language Models

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. In International Conference on Machine Learning (ICML) , 2024. URL https://arxiv.org/abs/2401.10020

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    G\"odel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement

    Xunjian Yin, Xinyi Wang, Liangming Pan, Li Lin, Xiaojun Wan, and William Yang Wang. Gödel agent: A self-referential agent framework for recursive self-improvement, 2024. URL https://arxiv.org/abs/2410.04444

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents

    Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin gödel machine: Open-ended evolution of self-improving agents, 2025. URL https://arxiv.org/abs/2505.22954

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  11. [11]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automated open-ended scientific discovery, 2024. URL https: //arxiv.org/abs/2408.06292

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    A survey of on-policy distillation for large language models,

    Mingyang Song and Mao Zheng. A survey of on-policy distillation for large language models,

    work page

  13. [13]

    URL https://arxiv.org/abs/2604.00626. 29

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Jiaqi Wei, Xiang Zhang, Yuejin Yang, Wenxuan Huang, Juntai Cao, Sheng Xu, Xiang Zhuang, Zhangyang Gao, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Chenyu You, Wanli Ouyang, and Siqi Sun. Unifying tree search algorithm and reward design for LLM reasoning: A survey, 2025. URL https://arxiv.org/abs/2510.09988

    work page arXiv 2025

  15. [15]

    Stop Hand-Holding Your Coding Agent: Engineering the Loops that Replace Step-by-Step Prompting

    Sandeco Macedo. Stop hand-holding your coding agent: Engineering the loops that replace step-by-step prompting, 2026. URL https://arxiv.org/abs/2607.00038

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Absolute Zero: Reinforced Self-play Reasoning with Zero Data

    Andrew Zhao, Yiran Wu, Yang Yue, Tong Wu, Quentin Xu, Yang Yue, Matthieu Lin, Shenzhi Wang, Qingyun Wu, Zilong Zheng, and Gao Huang. Absolute zero: Reinforced self-play reasoning with zero data, 2025. URL https://arxiv.org/abs/2505.03335

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    R-zero: Self-evolving reasoning LLM from zero data,

    Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, Ruosen Li, Jiaxin Huang, Haitao Mi, and Dong Yu. R-zero: Self-evolving reasoning LLM from zero data,

    work page

  18. [18]

    R-Zero: Self-Evolving Reasoning LLM from Zero Data

    URL https://arxiv.org/abs/2508.05004. ICLR 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems (NeurIPS) , 2023. URL https://arxiv.org/abs/2303.1 1366

    work page 2023

  20. [20]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    Charlie Snell, Jae-Hoon Lee, Kelvin Xu, and A viral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling model parameters, 2024. URL https://arxiv.or g/abs/2408.03314

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    SymbolicAI: A framework for logic-based approaches combining generative models and solvers

    Marius-Constantin Dinu, Claudiu Leoveanu-Condrei, Markus Holzleitner, Werner Zellinger, and Sepp Hochreiter. SymbolicAI: A framework for logic-based approaches combining gener- ative models and solvers, 2024. URL https://arxiv.org/abs/2402.00854

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework

    Jiajing Zhang, Jiamei Jiang, Chenyang Zhang, Feifei Mo, Linjing Li, and Daniel Zeng. To- wards reliable and robust LLM planning: Symbolic feedback-driven iterative self-refinement framework, 2026. URL https://arxiv.org/abs/2606.27757

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    SQL-o1: A Self-Reward Heuristic Dynamic Search Method for Text-to-SQL

    Shuai Lyu, Haoran Luo, Ripeng Li, Zhonghong Ou, Jiangfeng Sun, Yang Qin, Xiaoran Shang, Meina Song, and Yifan Zhu. SQL-o1: A self-reward heuristic dynamic search method for text- to-SQL, 2025. URL https://arxiv.org/abs/2502.11741

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Hallucination Detection-Guided Preference Optimization for Clinical Summarization

    Shamanth Kuthpadi Seethakantha, Dung Thai, Vara Prasad Gudi, Simran Tiwari, Rami Matar, A vijit Mitra, Wenlong Zhao, Wael Salloum, and Andrew McCallum. Hallucination detection-guided preference optimization for clinical summarization, 2026. URL https://arxi v.org/abs/2605.28910

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    LongSumEval: Question-Answering Based Evaluation and Feedback-Driven Refinement for Long Document Summarization

    Huyen Nguyen, Haoxuan Zhang, Yang Zhang, Haihua Chen, and Junhua Ding. LongSumEval: Question-answering based evaluation and feedback-driven refinement for long document sum- marization, 2026. URL https://arxiv.org/abs/2604.25130

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  26. [26]

    LLM-Personalize: Aligning LLM Planners with Human Preferences via Reinforced Self-Training for Housekeeping Robots

    Dongge Han, Trevor McInroe, Adam Jelley, Stefano V. Albrecht, Peter Bell, and Amos Storkey. LLM-personalize: Aligning LLM planners with human preferences via reinforced self-training for housekeeping robots, 2024. URL https://arxiv.org/abs/2404.14285

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    Adaptive self-improvement LLM agentic system for ML library development, 2025

    Genghan Zhang, Weixin Liang, Olivia Hsu, and Kunle Olukotun. Adaptive self-improvement LLM agentic system for ML library development, 2025. URL https://arxiv.org/abs/2502.0 2534. 30

    work page 2025

  28. [28]

    What Does LLM Refinement Actually Improve? A Systematic Study on Document-Level Literary Translation

    Shaomu Tan, Dawei Zhu, Ke Tran, Michael Denkowski, Sony Trenous, Bill Byrne, Leonardo Ribeiro, and Felix Hieber. What does LLM refinement actually improve? a systematic study on document-level literary translation, 2026. URL https://arxiv.org/abs/2605.13368

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [29]

    Large Language Models Cannot Self-Correct Reasoning Yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet, 2024. URL https://arxiv.org/abs/2310.01798. ICLR 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    When Does Intrinsic Self-Correction Help? A Task-Sensitive Analysis

    Elroy Stav, Dvir Berlowitz, Maayan Orner, and Sarit Kraus. When does intrinsic self- correction help? a task-sensitive analysis, 2026. URL https://arxiv.org/abs/2606.23196

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  31. [31]

    When are likely answers right? On Sequence Probability and Correctness in LLMs

    Johannes Zenn and Jonas Geiping. When are likely answers right? on sequence probability and correctness in LLMs, 2026. URL https://arxiv.org/abs/2606.27359

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  32. [32]

    Ask, Don't Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement

    Sangwoo Cho, Kushal Chawla, Pengshan Cai, Zefang Liu, Chenyang Zhu, Shi-Xiong Zhang, and Sambit Sahu. Ask, don’t judge: Binary questions for interpretable LLM evaluation and self-improvement, 2026. URL https://arxiv.org/abs/2606.27226

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    LL3M: Large Language 3D Modelers

    Shuang Lu, Chen Guan, Nam Anh Dinh, Itai Lang, Ari Holtzman, and Rana Hanocka. Ll3m: Large language 3d modelers, 2025. URL https://arxiv.org/abs/2508.08228

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    AgenticDB: Agentic Performance Reconfiguration for Database Workloads

    Xiaomin Yang, Chaozheng Wang, Chen Zheng, Heng Zhang, and Yanjun Wu. AgenticDB: Agentic performance reconfiguration for database workloads, 2026. URL https://arxiv.org/ abs/2606.20318

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  35. [35]

    AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

    Zepeng Li, Jie Ren, Zhanyong Tang, Jie Zheng, and Zheng Wang. AutoPass: Evidence-guided LLM agents for compiler performance tuning, 2026. URL https://arxiv.org/abs/2606.20373

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [36]

    LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks

    Po-Nien Kung, Linfeng Song, Dawsen Hwang, Jinsung Yoon, Chun-Liang Li, Simone Sev- erini, Mirek Olšák, Edward Lockhart, Quoc V Le, Burak Göktürk, Thang Luong, Tomas Pfister, and Nanyun Peng. LEAP: Supercharging LLMs for formal mathematics with agentic frameworks, 2026. URL https://arxiv.org/abs/2606.03303

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  37. [37]

    KVerus: Scalable and Resilient Formal Verification Proof Generation for Rust Code

    Yuwei Liu, Xinyi Wan, Yanhao Wang, Minghua Wang, Lin Huang, and Tao Wei. KVerus: Scalable and resilient formal verification proof generation for rust code, 2026. URL https: //arxiv.org/abs/2605.03822

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [38]

    KBSpec: LLM-driven Formal Specification Generation with Evolving Domain Knowledge Base

    Wenhan Wang and Zeyu Sun. KBSpec: LLM-driven formal specification generation with evolving domain knowledge base, 2026. URL https://arxiv.org/abs/2606.21339

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  39. [39]

    Verifier-Guided Code Translation via Meta-Step Decoding

    Tianyang Zhou, Somesh Jha, Mihai Christodorescu, Kirill Levchenko, and Varun Chan- drasekaran. Verifier-guided code translation via meta-step decoding, 2026. URL https: //arxiv.org/abs/2605.17626

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    What Drives Interactive Improvement from Feedback?

    Bartłomiej Cupiał, Jan Lojek, Mikołaj Garstecki, Szymon Pobłocki, Alicja Ziarko, and Piotr Miłoś. What drives interactive improvement from feedback?, 2026. URL https://arxiv.org/ abs/2606.30774

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  41. [41]

    Falsification, Not Exposure: An Internally Preregistered Placebo-Controlled Decomposition of Self-Repair Feedback in Frozen Small Code Models

    Mehmet İşcan. Falsification, not exposure: An internally preregistered placebo-controlled decomposition of self-repair feedback in frozen small code models, 2026. URL https://arxiv. org/abs/2606.31511

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  42. [42]

    Feedback Over Form: Why Execution Feedback Matters More Than Pipeline Topology in 1-3B Code Generation

    Charles Junichi McAndrews. Feedback over form: Why execution feedback matters more than pipeline topology in 1-3b code generation, 2026. URL https://arxiv.org/abs/2604.21950. 31

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [43]

    RubricRefine: Improving Tool-Use Agent Reliability with Training-Free Pre-Execution Refinement

    Will LeVine, Brendan Evers, Sam Saltwick, and Abhay Venkatesh. RubricRefine: Improving tool-use agent reliability with training-free pre-execution refinement, 2026. URL https://ar xiv.org/abs/2605.09730

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Unlocking LLM code correction with iterative feedback loops,

    Le Zhang and Suresh Kothari. Unlocking LLM code correction with iterative feedback loops,

    work page

  45. [45]

    URL https://arxiv.org/abs/2606.17514

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [46]

    Denoising Iterative Self-Correction: Structured Verification Loops for Reliable LLM Reasoning

    Shen Yin, David Ken, and Joel Stremmel. Denoising iterative self-correction: Structured verification loops for reliable LLM reasoning, 2026. URL https://arxiv.org/abs/2606.21724

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  47. [47]

    FLARE: Fine-Grained Diagnostic Feedback for LLM Code Refinement

    Yinsheng Yao, Hongxiang Zhang, Weixi Tong, and Tianyi Zhang. FLARE: Fine-grained diagnostic feedback for LLM code refinement, 2026. URL https://arxiv.org/abs/2606.03852

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  48. [48]

    CoSPlay: Cooperative Self-Play at Test-Time with Self-Generated Code and Unit Test

    Zhangyi Hu, Chenhui Liu, Tian Huang, Jindong Li, Yang Yang, Jiemin Wu, Zining Zhong, Menglin Yang, and Yutao Yue. CoSPlay: Cooperative self-play at test-time with self- generated code and unit test, 2026. URL https://arxiv.org/abs/2605.23491

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  49. [49]

    Kestrel: Grounding self-refinement for L VLM hallucination mitigation, 2026

    Jiawei Mao, Hardy Chen, Haoqin Tu, Yuhan Wang, Letian Zhang, Zeyu Zheng, Huaxiu Yao, Zirui Wang, Cihang Xie, and Yuyin Zhou. Kestrel: Grounding self-refinement for L VLM hallucination mitigation, 2026. URL https://arxiv.org/abs/2603.16664

    work page arXiv 2026

  50. [50]

    Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding

    Shuimu Chen, Yihan Chen, Yuanshen Guan, Zebang Cheng, Zeyu Zhang, Su-Juan Qin, Bin Xia, Jiaran Li, Wenming Yang, and Fei Ma. Reflect-r1: Evidence-driven reflection for self- correction in long video understanding, 2026. URL https://arxiv.org/abs/2606.27922

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  51. [51]

    FiRe: Fine-grained Multimodal Reasoning for Enhanced Image Generation

    Yong-Jin Kim, Yoonjin Oh, Yerin Kim, Hyomin Kim, Jeeyoung Yun, Yujung Heo, Minjun Kim, and Sungwoong Kim. FiRe: Fine-grained multimodal reasoning for enhanced image generation, 2026. URL https://arxiv.org/abs/2604.13491

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  52. [52]

    Proprio: Latent Self-Scoring and Inference-Time Refinement for Physically Plausible Video Generation

    Mariam Hassan, Kaouther Messaoud, Wuyang Li, and Alexandre Alahi. Proprio: Latent self- scoring and inference-time refinement for physically plausible video generation, 2026. URL https://arxiv.org/abs/2605.28230

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  53. [53]

    ActiveScope: Actively Seeking and Correcting Perception for MLLMs

    Yajing Wang, Chao Bi, Junshu Sun, Shufan Shen, Zhaobo Qi, Shuhui Wang, and Q Huang. ActiveScope: Actively seeking and correcting perception for MLLMs, 2026. URL https: //arxiv.org/abs/2606.24292

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  54. [54]

    Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks

    Yunqi Xue, Zhijiang Li, Philip Torr, and Jindong Gu. Safe autoregressive image generation with iterative self-improving codebooks, 2026. URL https://arxiv.org/abs/2606.27147

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  55. [55]

    Paying More Attention to Visual Tokens in Self-Evolving Large Multimodal Models

    Shravan Venkatraman, Ritesh Thawkar, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Salman Khan, and Fahad Khan. Paying more attention to visual tokens in self- evolving large multimodal models, 2026. URL https://arxiv.org/abs/2606.27373

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  56. [56]

    Personal Visual Context Learning in Large Multimodal Models

    Zihui Xue, Ami Baid, Sangho Kim, Mi Luo, and Kristen Grauman. Personal visual context learning in large multimodal models, 2026. URL https://arxiv.org/abs/2605.10936

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  57. [57]

    Each Judge Its Own Yardstick: Discovering Per-VLM Taxonomies for Physical Video Evaluation

    Yu Cao, Ziquan Liu, Zhensong Zhang, Jiankang Deng, Shaogang Gong, and Jifei Song. Each judge its own yardstick: Discovering per-VLM taxonomies for physical video evaluation, 2026. URL https://arxiv.org/abs/2606.22918

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  58. [58]

    Query-conditioned test-time self-training for large language models, 2026

    Chaehee Song, Minseok Seo, Yeeun Seong, Doyi Kim, and Changick Kim. Query-conditioned test-time self-training for large language models, 2026. URL https://arxiv.org/abs/2605.133 69. 32

    work page 2026

  59. [59]

    Continual Self-Improvement with Lightweight Experiential Latent Memories

    Vaggelis Dorovatas, Nancy Kalaj, and Rahaf Aljundi. Continual self-improvement with lightweight experiential latent memories, 2026. URL https://arxiv.org/abs/2606.17803

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  60. [60]

    Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories

    Ali Behrouz, Farnoosh Hashemi, and Vahab Mirrokni. Language models need sleep: Learning to self-modify and consolidate memories, 2026. URL https://arxiv.org/abs/2606.03979

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  61. [61]

    Truly Self-Improving Agents Require Intrinsic Metacognitive Learning

    Tennison Liu and Mihaela van der Schaar. Truly self-improving agents require intrinsic metacognitive learning, 2025. URL https://arxiv.org/abs/2506.05109

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  62. [62]

    Environment-Grounded Automated Prompt Optimization for LLM Game Agents

    Rean Clive Fernandes, Lukas Fehring, Theresa Eimer, Marius Lindauer, and Matthias Feurer. Environment-grounded automated prompt optimization for LLM game agents, 2026. URL https://arxiv.org/abs/2606.17838

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  63. [63]

    Yuxuan Wan, Tianqing Fang, Zaitang Li, Yintong Huo, Wenxuan Wang, Haitao Mi, Dong Yu, and Michael R. Lyu. Inference-time scaling of verification: Self-evolving deep research agents via test-time rubric-guided verification, 2026. URL https://arxiv.org/abs/2601.15808

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  64. [64]

    The Red Queen G\"odel Machine: Co-Evolving Agents and Their Evaluators

    Alex Iacob, Andrej Jovanović, William F. Shen, Daniel Burkhardt, Meghdad Kurmanji, Nurbek Tastan, Lorenzo Sani, Niccolò Alberto Elia Venanzi, Ambroise Odonnat, Zeyu Cao, Bill Marino, Xinchi Qiu, and Nicholas D. Lane. The red queen gödel machine: Co-evolving agents and their evaluators, 2026. URL https://arxiv.org/abs/2606.26294

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  65. [65]

    Sc.) Yang

    Ziyang Liu, Xinyan Guo, Xuchen Wei, Han Hao, and Liu (M. Sc.) Yang. Escher-loop: Mutual evolution by closed-loop self-referential optimization, 2026. URL https://arxiv.org/abs/2604 .23472

    work page 2026

  66. [66]

    QueenBee Planner: Skill-Evolving Communication Topologies for Token-Efficient LLM Multi-Agent Systems

    Congjia Tian, Yuhang Yao, and Jiaming Cui. QueenBee planner: Skill-evolving com- munication topologies for token-efficient LLM multi-agent systems, 2026. URL https: //arxiv.org/abs/2606.27492

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  67. [67]

    Learning from Failure: Inference-Time Self-Improvement for Computer-Use Agents

    Xueqiao Sun, Xiaohan Wang, Ludwig Schmidt, Serena Yeung-Levy, and Yuhui Zhang. Learn- ing from failure: Inference-time self-improvement for computer-use agents, 2026. URL https://arxiv.org/abs/2606.31270

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  68. [68]

    Gang Liao, Y He, Abdullah Ozturk, Zhouyang Li, Ying Wang, Zichang Guo, Hongsen Qin, Yaobin Qin, Tao Yang, Zewei Jiang, Dianshi Li, Jort Gemmeke, Jiangyuan Li, Liyuan Li, Nathan Yan, Masha Basmanova, Uladzimir Pashkevich, Matt Steiner, Pedro Pedreira, Rob Fergus, Anirudh Goyal, C E Wu, Gaoxiang Liu, Andrew Witten, and Daniel J. Abadi. Experience graphs: Th...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  69. [69]

    The meta-agent challenge: Are current agents capable of autonomous agent development?, 2026

    Xinyu Lu, Tianshu Wang, Pengbo Wang, Zujie Wen, Zhiqiang Zhang, Jun Zhou, Boxi Cao, Yaojie Lu, Hongyu Lin, Xianpei Han, and Le Sun. The meta-agent challenge: Are current agents capable of autonomous agent development?, 2026. URL https://arxiv.org/abs/2606.0 4455

    work page 2026

  70. [70]

    SAGE: A quantitative evaluation of socialized evolution in agent ecosystems, 2026

    Linyue Pan, Yaoming Zhu, Lin Qiu, Xuezhi Cao, and Xunliang Cai. SAGE: A quantitative evaluation of socialized evolution in agent ecosystems, 2026. URL https://arxiv.org/abs/26 06.03544

    work page 2026

  71. [71]

    Multi-Agent Reasoning Improves Compute Efficiency: Pareto-Optimal Test-Time Scaling

    Florian Valentin Wunderlich, Lars Benedikt Kaesberg, Jan Philip Wahle, Terry Ruas, and Bela Gipp. Multi-agent reasoning improves compute efficiency: Pareto-optimal test-time scaling, 2026. URL https://arxiv.org/abs/2605.01566. 33

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  72. [72]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models, 2023. URL https://arxiv.org/abs/2305.16291. TMLR 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  73. [73]

    SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-Refinement

    Srishti Gautam, Arjun Radhakrishna, and Sumit Gulwani. SkillAxe: Sharpening LLM- authored agent skills through evaluation-guided self-refinement, 2026. URL https://arxi v.org/abs/2606.10546

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  74. [74]

    Skill-R1: Agent Skill Evolution via Reinforcement Learning

    Yash Vishe, Rohan Surana, Xunyi Jiang, Zihan Huang, Xintong Li, Nikki Lijing Kuang, Tong Yu, Ryan A. Rossi, Jingbo Shang, Julian McAuley, and Junda Wu. Skill-r1: Agent skill evolution via reinforcement learning, 2026. URL https://arxiv.org/abs/2605.09359

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  75. [75]

    SkillRevise: Improving LLM-Authored Agent Skills via Trace-Conditioned Skill Revision

    Yuxuan Liu, Zhaochen Su, Lingyun Xie, Yuhao Zhang, Qing Zong, Jiahe Guo, Zhongwei Xie, Yiyan Ji, Yauwai Yim, Hongyu Luo, Xiyu Ren, Ruan Chenyu, Haoran Li, and Yangqiu Song. SkillRevise: Improving LLM-authored agent skills via trace-conditioned skill revision, 2026. URL https://arxiv.org/abs/2606.01139

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  76. [76]

    AlgoSkill: Learning to Design Algorithms by Scheduling Human-Like Skills

    Xinyuan Song, Zekun Cai, and L Zhao. AlgoSkill: Learning to design algorithms by scheduling human-like skills, 2026. URL https://arxiv.org/abs/2606.29999

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  77. [77]

    SkillMaster: Toward Autonomous Skill Mastery in LLM Agents

    Min Yang, Jinghua Piao, Xu Xia, Xiaochong Lan, Jiaju Chen, Yongshun Gong, and Yong Li. SkillMaster: Toward autonomous skill mastery in LLM agents, 2026. URL https://arxiv.or g/abs/2605.08693

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  78. [78]

    FederatedSkill: Federated Learning for Agentic Skill Evolution

    Jingbo Yang, Guanyu Yao, Yang Zhang, Ramana Rao Kompella, Gaowen Liu, and Shiyu Chang. FederatedSkill: Federated learning for agentic skill evolution, 2026. URL https: //arxiv.org/abs/2606.03143

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  79. [79]

    SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems

    Yangbo Wei, Zhen Huang, Shaoqiang Lu, Junhong Qian, Qifan Wang, Chen Wu, and Lei He. SkillSmith: Co-evolving skills and tools for self-improving agent systems, 2026. URL https://arxiv.org/abs/2606.01314

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  80. [80]

    Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills

    Chuan Xiao, Zhengbo Jiao, Shaobo Wang, Wei Wang, Bing Zhao, Hu Wei, Linfeng Zhang, and Lin Qu. Socratic-SWE: Self-evolving coding agents via trace-derived agent skills, 2026. URL https://arxiv.org/abs/2606.07412

    work page internal anchor Pith review Pith/arXiv arXiv 2026

Showing first 80 references.