Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Skill edits that do not commute can be screened with microsecond Lie-bracket tests, recovering full ordered-pair repair quality while cutting expensive model validations by nearly an order of magnitude.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 05:13 UTC pith:3NZIYCEX

load-bearing objection Real systems idea (screen noncommuting skill edits before LLM validation) wrapped in heavy algebroid language; speedups are concrete on controlled tasks but the screen mostly re-ranks planted dependency metadata. the 5 major comments →

arxiv 2607.11493 v1 pith:3NZIYCEX submitted 2026-07-13 cs.LG cs.AImath.CT

Agentic Skill Optimization over Lie Algebroids

Sridhar Mahadevan This is my paper
classification cs.LG cs.AImath.CT
keywords skill optimizationlarge language modelsLie algebroidstangent categoriesinfinitesimal causalityagentic systemsedit compositionvalidation screening
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Agentic systems improve by editing structured skills—prompts, schemas, plans, validators, and traces—but those edits are not independent coordinates. Distinct policies can look the same on the current document while carrying different routing or template state, and the order of two repairs often changes the outcome. This paper argues that the right local geometry is a controlled Lie algebroid over typed, anchored Markdown workflows: edit policies sit upstairs, an anchor maps each policy to its visible artifact effect, the kernel holds latent structure, and the algebroid bracket records order-sensitivity. Operationally, a cheap static bracket screen ranks which ordered pairs deserve costly rollout validation. On controlled and trace-derived benchmarks, that screen recovers the exhaustive ordered-pair solution while replacing most served large-model calls with microsecond probes, with a reported peak near 15× wall-clock speedup versus validating every pair through a large hosted model.

Core claim

When skill repairs are order-dependent, modeling available edit policies as sections of a controlled Lie algebroid over typed Markdown workflows lets a static noncommutation (bracket) screen select a small validation budget that still matches exhaustive ordered-pair search, because most of the combinatorial cost is replaced by microsecond algebraic probes before any large language model is run.

What carries the argument

LASKO: a controlled Lie algebroid A → Md with anchor ρ, so that sections are edit policies, ρ(s) is the visible Markdown or workflow effect, ker(ρ) is latent template/routing structure, and the algebroid bracket [s,t]_A screens noncommuting edit composition before expensive validation.

Load-bearing premise

The cheap, application-specific noncommutation proxies (shared anchors, read/write overlap, residual overlap, scenario-local dependencies) must actually rank the ordered pairs that matter for later validation, rather than only echoing structure already written into the benchmark graphs.

What would settle it

On a live skill-optimization task whose productive ordered edges are not already encoded in the same metadata the proxy reads, run LASKO’s bracket screen plus a fixed small validation budget against exhaustive ordered-pair validation with the same served model: if LASKO fails to recover the exhaustive score while random same-budget pairs do no better either, the screening claim fails for that setting.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Short-horizon single-edit skill loops will systematically miss productive ordered repair chains that a bracket screen can still surface within a linear validation budget.
  • Validation economics for self-editing agents can be reorganized into a two-layer queue: microsecond static noncommutation probes first, then focused served-model checks only on high-bracket pairs.
  • Hidden kernel directions (routing, templates, unfilled slots) become first-class objects to monitor when two edits leave the document looking the same but change future composition.
  • The same anchor/bracket accounting transfers from schematic ten-anchor chains to agent workflow contracts, financial research workflows, and predictive-state causal manifolds built from real run traces.

Where Pith is reading between the lines

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

  • If the proxy remains faithful outside hand-designed graphs, production skill optimizers could treat noncommuting pairs as a first-class search primitive rather than as an after-the-fact debugging story.
  • The kernel residual is a natural place to look for brittleness when a skill “works” on held-out scores yet still fails to compose with later edits or tools.
  • A natural next stress test is whether bracket screening still concentrates value when the productive order is discovered only from noisy live rollouts, not from a pre-written edge catalog.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes LASKO (Lie Algebroid SKill Optimization), framing agentic skill editing as optimization over a controlled Lie algebroid A → Md of edit policies above a tangent category of typed, anchored Markdown workflows. The anchor ρ maps controlled edit sections to visible artifact changes, ker(ρ) is interpreted as latent template/routing structure, and algebroid brackets are proposed as diagnostics of noncommuting edit composition. Empirically, the paper reports that inexpensive static “bracket” screens (microseconds) can prioritize ordered edit pairs so that a small number of expensive served-LLM validations recover the exhaustive ordered-pair score on controlled multi-anchor, agent-workflow, 10-K, Democritus, and PSR-LASKO tasks, with a reported peak wall-clock speedup of ~14.85× versus brute-force validation on DeepSeek V3.1 4-bit for a WaPo coffee PSR-LASKO run.

Significance. If the geometric and empirical claims hold beyond hand-designed graphs, the paper would supply a useful validation-economics primitive for self-improving agents: treat noncommuting skill repairs as first-class objects and spend expensive rollouts only on high-interaction pairs. Strengths that should be credited include (i) a clear operational separation of cheap static probes from served validation (Appendix A), (ii) multi-model wall-clock accounting rather than only probe counts, and (iii) an explicit attempt to connect skill optimization to Lie-algebroid and tangent-category structure rather than flat prompt search. Even if the full algebroid story is only partially realized, the prioritization-of-ordered-repairs idea is practically relevant to SKILLOPT-style systems.

major comments (5)
  1. §10.3–10.5 and Appendix A.1: the central speedup claim is load-bearing on application-specific static noncommutation proxies (read/write anchor overlap, residual-component overlap, scenario-local ordered dependencies, etc.). In the ten-anchor, agent-workflow, and 10-K benchmarks, productive ordered pairs are hand-specified (e.g., schema→normalization, table extraction before numeric gate), and the proxy is built from the same anchor/read-write and scenario-dependency metadata that encodes those edges. Later Democritus/PSR cases reuse the same architecture. As written, the screen largely re-ranks a known noncommutative edit graph rather than independently estimating an algebroid bracket [s,t]_A or discovering order-sensitivity without planted structure. Please either (a) provide a benchmark where productive order is not encoded in the proxy features, or (b) reframe the empirical claim as
  2. Abstract and §1 vs Appendix A.1: the abstract states that LASKO “substitutes inexpensive Lie-bracket screening tests that run in microseconds.” Appendix A.1 then defines the operational object as an application-specific static noncommutation proxy, not a computed Lie algebroid bracket on sections of A. This is a load-bearing terminology gap: readers will take “Lie-bracket screening” as evidence for the geometric formalism, whereas the measured object is a hand-crafted compatibility score. Align the abstract, introduction, and experimental claims with what is actually computed, or report an actual bracket estimator (order-contrast of finite differences / residual vectors) that does not hard-code the answer graph.
  3. Proposition 1 (§6) and the experimental base: Proposition 1 is explicitly conditional on edit fibers being modules and rewrites having compatible pushforwards. The experiments never verify these hypotheses for the Markdown/workflow objects used (skill cards, Democritus traces, PSR manifolds). Without that, the claim that the base is a tangent category—and therefore that involution-algebroid closure is the right workflow condition—remains formal scaffolding rather than an established property of the systems under study. Either supply a concrete verification for one experimental base, or demote the tangent-category claim to a modeling hypothesis and state what would falsify it.
  4. §9 Conjecture 1 and the LASKO loss: the combined loss L_LASKO with weights λ_IC, λ_ker, λ_DB, λ_A and the low-curvature fixed-point conjecture are presented as the optimization principle, but the reported experiments do not optimize this loss; they rank ordered pairs by a static proxy and validate. The free parameters (top-k after screen, λ weights, proxy scoring weights) are not ablated. For the central claim to rest on LASKO-as-optimization rather than LASKO-as-prioritizer, show that the residual terms (R_Md, R_ker, R_DB) predict repair stability on held-out or non-planted tasks, or clearly separate the prioritization algorithm from the untested optimality conjecture.
  5. Native baselines (§10.8–10.9): SearchQA and ALFWorld are correctly labeled micro-benchmarks, but they do not yet support the abstract’s order-of-magnitude claim. SearchQA shows a gated accept/reject with a suggestive fiber-level soft-score signal; ALFWorld closes one ordered repair edge with a hand-written navigation rule. Neither demonstrates that algebroid screening reduces validation cost relative to SKILLOPT on a live environment. Either expand these to a quantitative LASKO-vs-exhaustive comparison, or confine the speedup claim strictly to the controlled/PSR served-validation tables and state the live-environment status as preliminary.
minor comments (6)
  1. Figure 1 and Figure 2 are helpful but dense; label ρ, ker(ρ), and [s,t]_A more explicitly in the figure captions so the geometric vocabulary is self-contained without §3.
  2. Notation drift: Md, WfMd, TWfMd, A_psr, and M_psr appear with slightly different roles; a single notation table early in §6–7 would help.
  3. Several related works are the author’s own 2026 manuscripts (IC, BRIDGE/SKFM, Categories for AGI, Odyssey). Brief one-sentence contrasts of what is new here versus those would help non-specialist readers.
  4. Typographical/formatting issues: title casing in headers (“AGENTICSKILLOPTIMIZATION”), missing spaces in compound names (SKILLOPT, LASKO), and occasional double spaces in the preprint body.
  5. Table-like method comparisons in §10.3–10.7 would be clearer as numbered tables with standard captions rather than inline monospace blocks.
  6. Clarify whether the DeepSeek V3.1 “671B parameters” figure in the abstract is the full model size or the served 4-bit local instance actually timed; the appendix table is the authoritative source and should be cross-referenced.

Circularity Check

4 steps flagged

Empirical speedup is largely by construction: productive ordered edges are planted in the benchmarks, and the 'Lie-bracket' screen ranks pairs from the same read/write and scenario-dependency metadata that defines those edges.

specific steps
  1. self definitional [§10.3 Ten-anchor controlled benchmark; §10.4 Agent-workflow controlled benchmark]
    "Five ordered pairs are productive: schema→normalization, evidence→citation, abstention→format, tool contract→routing, validator→final answer. ... The LASKO-prioritized method first computes a cheap bracket proxy from the anchor/order structure and then validates only the ten highest-bracket ordered pairs. ... The bracket proxy is computed from cheap workflow metadata. Each edit section has read and write anchors; a pair receives high bracket score when the first edit writes an anchor read by the second, when the reverse order is also plausible, or when the pair appears as a scenario-local orde"

    Productive ordered pairs are defined by construction as specific ordered dependencies on the anchor chain / scenario graph. The 'bracket' score is then defined to be high precisely when a pair exhibits that same read/write or scenario-local ordered dependency. Recovering exhaustive score 1.000 by validating the top-bracket pairs is therefore verifying the planted graph by reading the planted graph, not an independent measurement of noncommuting edit composition.

  2. fitted input called prediction [Appendix A.1 Application-specific bracket proxies; §10.7 PSR-LASKO; abstract speedup claim]
    "Operationally, however, the proxy used to rank section pairs depends on the artifact being optimized. In PSR-LASKO, the proxy combines read/write dependencies between PSR sections, overlap in residual components (RH, Rres, RW, Rcausal, Rscope, RDB), context-chart interactions, witness dependencies, Hankel row/column effects, and drift pressure induced by off-query clusters. ... Thus the invariant idea is not a single universal formula for the cheap probe. It is the architecture: application-specific static noncommutation proxy ⇒ focused expensive validation. ... LASKO achieved a speedup of alm"

    The abstract presents microsecond 'Lie-bracket screening' as the reason expensive LLM validations can be cut by ~15×. Operationally the probe is an application-specific static ranking built from the same residual families, read/write fields, and ordered-edge catalog later used as the validation target. Selecting high-proxy pairs and then reporting that those pairs close the catalog edges is a forced recovery of the input dependency structure, not a prediction of noncommutation outside that structure.

  3. renaming known result [§10.3–10.5 method tables; Appendix A architecture statement]
    "Method Bracket probes Validation probes Score ... LASKO-prioritized 90 10 1.000 ... Exhaustive ordered pairs 0 90 1.000 ... application-specific static noncommutation proxy ⇒ focused expensive validation."

    What is demonstrated is dependency-aware prioritization of a known noncommutative edit graph (validate high-dependency ordered pairs first), a standard engineering pattern. The paper renames that static prioritization as 'Lie-algebroid bracket screening' and presents the resulting validation reduction as evidence for algebroid geometry of skill edits, without an independent bracket estimator that would discover order-sensitivity when the answer is not already in the metadata.

  4. self citation load bearing [§1–2 geometric lineage; citations Mahadevan 2026b,d,c]
    "Infinitesimal Causality (IC) (Mahadevan, 2026d) supplies the first piece of this algebra. ... Earlier work proposed Lie-algebraic methods for latent-confounded causal discovery, specifically BRIDGE ... and SKFM ... (Mahadevan, 2026b). ... The motivating agent-system examples are BASKET and ROCKET, introduced in the broader diagrammatic agent-systems program of Categories for AGI (Mahadevan, 2026c)."

    The claim that skill optimization 'needs' controlled Lie algebroids, anchors ρ, ker(ρ), and bracket residuals as the right geometry is justified almost entirely by concurrent manuscripts by the same author (IC, BRIDGE/SKFM, Categories for AGI). Those citations are not machine-checked or externally fixed uniqueness results; they are the same research program. The empirical tables do not independently establish that algebroid structure (as opposed to ordinary dependency metadata) is what buys the speedup.

full rationale

The paper's central empirical claim—that microsecond 'Lie-bracket' probes recover exhaustive ordered-pair score 1.000 while cutting expensive served validations by an order of magnitude (peak ~14.85× on DeepSeek V3.1 4-bit)—is not an independent first-principles prediction of noncommutation. In the controlled suites that carry the claim (§10.3–10.5 and the later Democritus/PSR graphs that reuse the same architecture), productive ordered pairs are hand-specified, and Appendix A.1 / §10.4 define the cheap proxy from exactly the anchor read/write overlap and 'scenario-local ordered dependency' structure that encodes those pairs. Validating the top-scoring pairs therefore re-reads a planted dependency graph rather than estimating an algebroid bracket [s,t]_A or ker(ρ) from rollout dynamics. The geometric vocabulary (IC, BRIDGE/SKFM, Categories for AGI) is load-bearing only via concurrent self-citations and does not independently force the validation economics. SearchQA/ALFWorld/Odyssey traces are smaller diagnostics and do not rescue the main speedup tables. This is partial circularity of the empirical core (score 7), not a total definitional collapse of every claim: the paper does honestly label the suites as controlled and does run real served validators on selected sequences.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 5 invented entities

The load-bearing content is mostly modeling choices and constructed evaluation graphs, not fitted physical constants. The central empirical claim depends on treating skill artifacts as a tangent category, edit policies as algebroid sections, and static dependency proxies as brackets. Free parameters are mainly screening budgets and un-tuned multi-term loss weights. Invented entities are the LASKO stack and residual diagnostics; independent evidence outside this paper is thin because benchmarks are author-constructed and code is not shipped.

free parameters (3)
  • validation_budget_after_screen (top-k / top-2k pairs, focused final chunks) = task-specific (e.g. 10, 17, 19, 21, 2k)
    Number of expensive ordered-pair validations retained after bracket ranking is chosen by hand per benchmark (e.g., 10 of 90; 19 of 73; linear 2k scaling) and directly determines the reported speedup.
  • LLASKO weights λ_IC, λ_ker, λ_DB, λ_A = unspecified / not systematically fit
    Multi-term causal-curvature loss in §9 introduces free tradeoff weights; experiments primarily use ranking proxies rather than a fitted joint optimization of these λ's.
  • application-specific bracket proxy scoring weights = unspecified heuristic combination
    Appendix A.1 combines read/write dependencies, residual overlaps, chart/witness/Hankel effects, and drift pressure without a unique derived formula; relative weights are design choices.
axioms (6)
  • ad hoc to paper Typed anchored Markdown / workflow objects form a tangent category once edit fibers are modules and rewrites have compatible pushforwards (Proposition 1).
    Stated as conditional; raw Markdown is admitted not to suffice. Central formal base for LASKO.
  • domain assumption Available skill-edit policies form (or are well approximated by) sections of a Lie algebroid A→WfMd with anchor ρ preserving brackets.
    Definition 1 imported from Lie algebroid theory and imposed on agent edits; not derived from agent data.
  • ad hoc to paper Static noncommutation proxies estimate the pairs with large algebroid brackets well enough to replace most semantic validations.
    Operational heart of the speedup claim; Appendix A explicitly makes proxies application-specific rather than canonical brackets.
  • ad hoc to paper Low-curvature fixed points of the forward-backward repair operator are locally optimal stable skills (Conjecture 1).
    Unproved optimality principle used to interpret residuals as optimization geometry.
  • standard math Standard Lie algebroid Leibniz rule, anchor-bracket intertwining, and involutive-closure ideas from Mackenzie/Crainic-Fernandes/Burke-MacAdam apply after the Markdown lift.
    Background differential geometry/category theory cited and reused.
  • domain assumption SkillOpt-style propose/rollout/validate loops and BASKET/ROCKET/Democritus workflow stages are the right empirical substrate for testing the geometry.
    Systems vocabulary from Yang et al. 2026 and the author's Categories for AGI / Odyssey program.
invented entities (5)
  • LASKO (Lie Algebroid SKill Optimization) no independent evidence
    purpose: Name the optimization problem of editing typed skills under anchor, kernel, and bracket structure.
    Primary framework introduced by the paper; evidence is internal benchmarks, not external independent confirmation.
  • Typed Markdown tangent category Md / workflow category WfMd no independent evidence
    purpose: Provide a base 'manifold' of anchored artifacts and partially ordered workflows for tangent and reverse-tangent structure.
    Constructed formal objects; Proposition 1 is conditional and not machine-checked.
  • Controlled intervention algebroid of skill-edit policies with ker(ρ) latent procedural state no independent evidence
    purpose: Separate controlled edit modes from visible Markdown effects and host hidden routing/template structure.
    Interpretive lift of Lie algebroids; kernel is mostly proxied by trace disagreement when artifacts match.
  • LASKO residuals RMd, Rker, RDB and diagrammatic backpropagation curvature CurvD no independent evidence
    purpose: Quantify visible nonclosure, hidden interaction, and path-dependent blame for screening and loss design.
    Defined for the framework; empirical use often collapses to artifact/blame/trace distances on constructed graphs.
  • PSR-LASKO causal manifold Mpsr and repair sections no independent evidence
    purpose: Treat Democritus predictive-state outputs as base objects for ordered manifold repairs.
    Specialized invented substrate for later experiments; depends on author's Odyssey/Democritus stack.

pith-pipeline@v1.1.0-grok45 · 24260 in / 4695 out tokens · 55996 ms · 2026-07-14T05:13:23.539677+00:00 · methodology

0 comments
read the original abstract

Agentic systems increasingly improve themselves by editing skills: prompts, rubrics, plans, tool contracts, examples, validators, and traces. Skill edits are not independent coordinates in a vector space: they are local repairs to structured artifacts whose effects are observed only after rollout, validation, and critique. Distinct edits can have the same immediate visible effect while differing in routing context, template state, guardrail scope, or future composability. The order of edits can matter as well: repairing a schema before a normalization rule need not be equivalent to applying the same edits in the reverse order. This paper introduces a new framework for skill optimization called LASKO, for Lie Algebroid SKill Optimization. LASKO models typed, anchored Markdown skills as the base category and available edit policies as sections of a controlled Lie algebroid with anchor $\rho$. The anchor maps an edit policy to its visible Markdown effect; the kernel $\ker(\rho)$ represents latent template, routing, or implementation structure; and the algebroid bracket measures noncommuting edit composition. As shown in the paper, LASKO achieves order-of-magnitude speedups in skill optimization in our preliminary benchmark results, primarily because it substitutes inexpensive Lie-bracket screening tests that run in microseconds, before investing in expensive validations that require running large language models. On a causal extraction from natural language task, LASKO achieved a speedup of almost $15 \times$ compared to a brute-force approach that validated all edits by running them through a DeepSeek V3.1 4-bit model with 671B parameters.

Figures

Figures reproduced from arXiv: 2607.11493 by Sridhar Mahadevan.

Figure 1
Figure 1. Figure 1: Agentic skill optimization as algebroidic infinitesimal causality over typed Markdown artifacts. The [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The geometric terms used in the paper. The base space [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Learning in Infinitesimal Non-Compositional Sketches

    cs.LG 2026-07 conditional novelty 5.0

    The paper defines infinitesimal non-compositionality as the tangent-lift of factorization failures in learning sketches, and proposes learning as converging to a final coalgebra of iterated tangent lifts.

Reference graph

Works this paper leans on

14 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Matthew Burke and Rory B. B. MacAdam. Involution algebroids: a generalisation of Lie algebroids for tangent categories, 2019. URL https://arxiv.org/abs/1904.06594

    Pith/arXiv arXiv 2019

  2. [2]

    J. R. B. Cockett and G. S. H. Cruttwell. Differential structure, tangent structure, and SDG . Applied Categorical Structures, 22 0 (2): 0 331--417, 2014

    2014

  3. [3]

    Integrability of lie brackets

    Marius Crainic and Rui Loja Fernandes. Integrability of lie brackets. Annals of Mathematics, 157 0 (2): 0 575--620, 2003

    2003

  4. [4]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models. International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  5. [5]

    Kirill C. H. Mackenzie. General Theory of Lie Groupoids and Lie Algebroids, volume 213 of London Mathematical Society Lecture Note Series. Cambridge University Press, 2005

    2005

  6. [6]

    ALLORA : A lie-algebraic LoRA method for composable neural adapters, 2026 a

    Sridhar Mahadevan. ALLORA : A lie-algebraic LoRA method for composable neural adapters, 2026 a . Manuscript

    2026

  7. [7]

    Latent-confounded causal discovery via lie-bracket geometry, 2026 b

    Sridhar Mahadevan. Latent-confounded causal discovery via lie-bracket geometry, 2026 b . URL https://arxiv.org/abs/2606.19610

    Pith/arXiv arXiv 2026

  8. [8]

    Categories for AGI , 2026 c

    Sridhar Mahadevan. Categories for AGI , 2026 c . URL https://people.cs.umass.edu/ mahadeva/papers/catagi.pdf. Textbook manuscript

    2026

  9. [9]

    Infinitesimal causality, 2026 d

    Sridhar Mahadevan. Infinitesimal causality, 2026 d . URL https://arxiv.org/abs/2606.24621

    Pith/arXiv arXiv 2026

  10. [10]

    Odyssey: Constructing verifiable local truth-preserving foundation models, 2026 e

    Sridhar Mahadevan. Odyssey: Constructing verifiable local truth-preserving foundation models, 2026 e . URL https://arxiv.org/abs/2606.27593

    Pith/arXiv arXiv 2026

  11. [11]

    Causality: Models, Reasoning, and Inference

    Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2 edition, 2009

    2009

  12. [12]

    Donald B. Rubin. Causal inference using potential outcomes: Design, modeling, decisions. Journal of the American Statistical Association, 100 0 (469): 0 322--331, 2005

    2005

  13. [13]

    Causation, Prediction, and Search

    Peter Spirtes, Clark Glymour, and Richard Scheines. Causation, Prediction, and Search. MIT Press, 2 edition, 2000

    2000

  14. [14]

    SkillOpt : Executive strategy for self-evolving agent skills, 2026

    Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, and Chong Luo. SkillOpt : Executive strategy for self-evolving agent skills, 2026. URL https://arxiv.org/abs/2605.23904

    Pith/arXiv arXiv 2026