Pith. sign in

REVIEW 3 major objections 5 minor 63 references

SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A skill library can be compressed threefold by rewriting repeated routines as reversible macros, without losing the dependencies and verifier checks execution needs.

desk verdict A genuinely new section-level compression framework with strong end-task results, but the contract-preservation guarantee is conditional on an unvalidated extraction step. read the letter →

arxiv 2608.05604 v1 pith:5NM4GHDI submitted 2026-08-06 cs.CL cs.AI

Xingyu Tan , Xiaoyang Wang , Qing Liu , Xiwei Xu , Xin Yuan , Liming Zhu , Wenjie Zhang This is my paper · ORCID
classification cs.CLcs.AI
keywords LLMagentsagentskillsproceduralmemoryskillretrievalgraphcompressioncontractpreservationmacrorewritinglibraries
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Large language model agents increasingly rely on skill libraries, but the library unit is mismatched: skills are stored as full packages, compressed as text, and only turned into execution graphs after retrieval. SkillZip argues that compression should happen over contract-bearing section-level graphs, so that a repeated routine is rewritten as a reversible macro only when its interface, dependencies, and verifier hooks survive the rewrite. The paper reports a 3.46x compression ratio with 99.2% dependency preservation and 98.7% verifier reachability, and task gains of up to 12.2 points over the strongest baseline. If this holds, large skill libraries can expose shared sub-skill routines under tight context budgets while remaining safe to execute and expandable back to source.

What carries the argument

The load-bearing object is the ported macro node $M_g : I_g \Rightarrow O_g$: a nonterminal node that replaces a connected subgraph $g$ of section nodes, exposes typed boundary ports $I_g$ and $O_g$, and carries an expansion rule $M_g \Rightarrow (V_g, E_g, \chi_g)$ back to the original occurrence-specific sections. A motif is admitted as a macro only after four checks: BoundaryClear (the ports match the external edges), SignatureStable (I/O, roles, and resources agree across occurrences), DependencyClosed (cut dependencies become macro ports), and VerifierReachable (every state-changing operation keeps a verifier inside the macro or reachable from its output). The machinery's work is captured in Proposition 1, compositional structural lifting: for pairwise non-conflicting occurrences, simultaneous macro rewriting and source expansion restore the raw subgraph up to isomorphism, preserving typed dependencies and operation-to-verifier reachability. That is what connects compression to execution: a shorter context is trusted only because the missing detail is recorded and recoverable.

What would settle it

Take two textually similar routines that differ only in an implicit verifier hook stated in prose rather than in a test file or explicit assertion, run the contract extractor on both, and check whether the macro acceptance checks still keep them separate. The paper's own data predict the failure boundary: at 40% synthetic contract corruption, verifier reachability falls from 98.7% to 77.2% and task reward from 33.3 to 24.1, so a direct test would measure the same degradation when the corruption comes from extraction-model errors rather than synthetic field edits, and would look for any hydrated context in which a state-changing operation executes with no reachable verifier.

Watch

Extended reading notes

Core claim

The central claim is that contract-preserving compression is the right abstraction for agent skill libraries. SkillZip represents each skill not as a document but as a procedural graph whose nodes are typed sections (intent, input, precondition, operation, resource, failure, verifier, output), and compresses by mining recurring section-level motifs and replacing them with ported macro nodes, subject to boundary, signature, dependency-closure, and verifier-reachability checks. A macro is a reversible rewrite rule with an explicit expansion back to source sections, so the compressed library remains a faithful index of the raw library. At query time, the hydration component compiles a compact dependency-closed subgraph and expands macros only to the level the task needs. The end-to-end result, per the paper, is that SkillZip outperforms the strongest baselines on both technical and embodied agent benchmarks (up to 12.2 points) while compressing the active representation by 3.46x and keeping structural fidelity near the raw graph, at 99.2% dependency preservation and 98.7% verifier reachability.

Load-bearing premise

Everything rests on the extraction step: if Sec2Graph misses a real dependency, precondition, guard, or verifier hook when it opens a skill package into a graph, the macro acceptance checks validate the wrong graph and the contract-preserving property stops protecting execution, a failure mode the paper's own corruption runs exhibit (reward dropping from 33.3 to 24.1 and verifier reachability to 77.2 at 40% noise).

Editorial extensions

If this is right

  • Section-level retrieval replaces whole-package loading: the correct source skill moves to the top of the ranking, and the retrieval advantage over the strongest baseline widens from 6.2 points at 200 skills to 23.3 points at 100K skills.
  • Persistent compressed graphs give roughly a threefold reduction in active representation (3.46x, a 71.0% storage cut) without text-level lossiness, because macros are reversible rewrites rather than opaque summaries.
  • Compression and execution are coupled: dropping the dependency-closure or verifier checks costs 14 to 22 points of structural fidelity and several points of task reward, so "shorter" and "safe to run" are separable properties.
  • Execution evidence can maintain the library: the incremental maintenance component promotes recurring residuals into macros and demotes or splits macros whose verifier failure rates rise, at 0.22x the cost of periodic full recompression.
  • The approach transfers across six LLM backbones and both benchmark families without fine-tuning, which the paper reads as evidence that the benefit comes from changing the retrieval unit rather than from model strength.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If section-level contract-preserving compression scales the way the 100K-skill run suggests, large skill libraries could be shipped once as compressed procedural graphs, making retrieval a graph query over reusable routines rather than a semantic search over whole documents.
  • The reversible-macro discipline has a direct analogue in compiler intermediate representations: the same boundary, dependency-closure, and verifier checks could apply to compressing tool-use trajectories into skill libraries, where the verifier is a check that a tool call or state change actually succeeded.
  • The paper's own robustness numbers (precondition and guard extraction at 79.6 exact match, reward falling from 33.3 to 24.1 at 40% contract corruption) imply the bottleneck is upstream extraction quality rather than the compression machinery; a testable extension would measure how macro acceptance rates change when extraction quality is the controlled variable.
  • A natural stress test is adversarial: construct skills whose verifiers are implied only by prose warnings rather than explicit check fields, and ask whether the dependency-closure guarantee can silently detach an operation from its safety check, which is exactly the rejection case the paper's workbook-versus-CSV example is designed to catch.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SkillZip, a pipeline that converts agent skill packages into a section-level procedural graph (Sec2Graph), compresses recurring contract-valid motifs into reversible ported macros (MotifZip), retrieves and hydrates budgeted executable contexts from the compressed graph (PathHydrate), and incrementally updates the library from new skills and execution traces (ReZip). The central claim is that this representation preserves procedural contracts—boundary signatures, dependency closure, verifier reachability, and source-level expandability—while cutting context roughly three-fold, and that this improves end-task performance and retrieval quality across SkillsBench and ALFWorld. On the main benchmarks the best comparable configuration improves over SkillDAG by 6.0–12.2 points depending on backbone and benchmark, with a 3.46x compression ratio, 99.2% dependency preservation, and 98.7% verifier reachability; scaling experiments extend retrieval to 100K skills.

Significance. If the empirical claims hold, section-level procedural abstraction is a useful contribution to skill-library management: it addresses a real granularity mismatch and provides a concrete mechanism for reversible compression with verifier-aware hydration. The empirical package is unusually careful in several respects: repeated-run paired bootstrap confidence intervals and permutation p-values (Table 13), controlled corruption analysis (Table 5), component-level ablations (Table 3), active-storage and downstream-inflation accounting (Table 4), and scaling to 100K skills with measured latency (Table 6). The paper is also honest about the limits of its formal guarantee: Proposition 1 is explicitly structural rather than semantic. The weakness is that the structural metrics largely verify preservation of the contracts that the pipeline itself extracts; because extraction is imperfect (84.6 exact match in Appendix B.1), the headline fidelity numbers do not fully establish executability with respect to the original skill packages.

major comments (3)
  1. [Section 4.2, Algorithm 6, Appendix B.1] The load-bearing claim that SkillZip compresses while preserving procedural contracts is only established relative to the contracts that Sec2Graph itself extracts. DPR and VR in Algorithm 6 compare the hydrated context against the raw graph G and the hydration log L_q, but both G and G_zip are produced by the same extraction step; a dependency or verifier that Sec2Graph misses is absent from both and therefore cannot lower DPR/VR. Proposition 1 is explicit that the guarantee covers recorded interfaces, dependencies, verifier paths, and provenance, not unrecorded behavior or verifier correctness. Since Appendix B.1 reports 91.6 macro-F1 and 84.6 exact match for extraction, and Table 5 shows that 40% synthetic contract corruption reduces task reward from 33.3 to 24.1 and verifier reachability to 77.2, the headline 3.46x/99.2%/98.7% numbers do not by themselves show that compressed contexts preserve executable structure relative to the original skill packages. I ask for an oracle-based evaluation—for example, a gold contract graph built from human annotations, or a comparison of end-task reward under Sec2Graph-extracted versus oracle contracts—so the conditional nature of the guarantee is quantified rather than only acknowledged.
  2. [Section 4.2, Proposition 1] Proposition 1 is the only formal support for the reversible-expansion guarantee, but it is presented as a proof sketch and its assumptions are not verified for the subgraphs that PathHydrate actually retrieves. The proof assumes that the occurrence map and port map identify every internal node and every boundary-crossing edge, and that the query subgraph P contains either all or none of the internal nodes of each occurrence. It is not shown that the subgraphs P_q produced by Algorithm 4 always satisfy this all-or-none condition, especially after scaffold repair and partial macro hydration. Please state and prove the lifting invariant at the level of hydrated contexts, or show by construction that every P_q satisfies the condition.
  3. [Equations (3) and (7), Section 5] The main results depend on hand-set objective weights: alpha, lambda, and mu for MotifZip's compression gain and eta, beta, gamma, and delta for PathHydrate's subgraph objective. The paper reports no sensitivity analysis over these coefficients; Figure 4 varies only the token budget. Because the reported end-task advantage over SkillDAG is as large as 12.2 points, a small perturbation study or a grid over the weight vectors is needed to show that the gains and the compression/fidelity trade-off are not an artifact of the chosen coefficients. At minimum, report the range of weights over which the qualitative conclusions in Tables 1 and 2 are unchanged.
minor comments (5)
  1. [Definition 2] The relation name 'Reqires' should be spelled 'Requires'.
  2. [Section 3, Definition 4] The notion of a subgraph 'port' is used before it is formally defined; please define I_g and O_g and the port map phi_omega near Definition 4.
  3. [Table 2 and Appendix D] The Recover. metric is used in the main text but defined only in Appendix D; a one-sentence definition in the main text would help the reader interpret Table 2.
  4. [Appendix B.1] The synthetic corruption protocol removes or replaces contract fields at random, but real extraction errors are likely systematic (for example, implicit preconditions); a brief discussion of this limitation and, ideally, an error-type breakdown would strengthen the robustness claim.
  5. [Section 4.2, Equation (3)] The Risk(g) term is said to penalize weak verifier support, but it is evaluated only after the hard validity checks; please clarify whether the risk term can ever change the acceptance decision or only ranks already accepted motifs.

Circularity Check

1 steps flagged · score 6.0 of 10

DPR/VR structural-fidelity claims reduce to SkillZip's own acceptance checks; end-task reward remains externally grounded.

  1. self definitional [Section 4.3 (Closed-loop invariants), Section 5.2 (RQ3), Appendix A.6 (Algorithm 6).]
    "The same quantities are used by the structure-aware evaluation protocol, so the empirical metrics test whether compression remains executable rather than merely shorter."

    MotifZip's contract validation (Algorithm 3) accepts a motif only after DependencyClosed(g, chi_g, G) and VerifierReachable(g, chi_g, G) return true, and PathHydrate (Algorithm 4) performs AttachReachableVerifiers and RepairClosure. Algorithm 6 then computes DPR = EvalDepPreserve(L_q, G) and VR = EvalVerifierReach(L_q, G_zip) from the hydration log. The reported 99.2% DPR and 98.7% VR therefore measure whether the pipeline enforced its own acceptance predicates on the graph that Sec2Graph itself extracted; a dependency or verifier that extraction missed is absent from G and cannot lower the scores. The paper concedes the guarantee is 'structural rather than semantic' and does not establish equivalence of unrecorded behavior or verifier correctness.

full rationale

The central end-task claims are not circular: SkillsBench reward and ALFWorld success come from external verifiers and environments, and they are not fitted from SkillZip's internal weights (the α, λ, μ and η, β, γ, δ appear hand-set, not tuned to the reported test outcomes). The compression ratio and token reductions are also independent measurements. However, the headline 'contract-preserving' structural metrics (DPR, VR, and the related recovery rates) are computed from the same data structures that the system's own acceptance checks populate: MotifZip refuses to compress a motif unless DependencyClosed and VerifierReachable pass, and PathHydrate performs scaffold repair to restore those same properties. Algorithm 6 then reads those logged properties back out as measured outcomes. This makes the 99.2%/98.7% numbers true statements about SkillZip's self-consistency, not about execution safety; the paper honestly labels Proposition 1 as 'structural rather than semantic.' The corruption experiments in Appendix B.1 show that when the extracted contracts themselves are wrong, the guarantees degrade sharply, which confirms that the preservation guarantee inherits every error of Sec2Graph. The correct score is therefore 6: some structural claims reduce by construction to the system's own validators, but the primary end-task evaluation is independent and non-circular.

Assumptions & free parameters 7 free parameters · 4 assumptions · 2 invented entities

The central claim rests on hand-set weights and thresholds that are not fitted to external benchmarks, plus domain assumptions about the adequacy of LLM-based contract extraction and about graph-level contract preservation being sufficient for executable reuse. The framework introduces macro nodes and canonical prototypes as internal representations without independent falsifiable handles. These ledger entries bound the contribution: the empirical gains are externally validated, but the structural fidelity metrics are partly guaranteed by the algorithm's own acceptance checks.

free parameters (7)
  • MotifZip gain weights (alpha, lambda, mu) = alpha=0.5, lambda=0.3, mu=0.2
    Hand-set in Appendix D to balance reuse, boundary cut, and risk; no sensitivity analysis reported.
  • PathHydrate objective weights (eta, beta, gamma, delta) = eta=0.4, beta=gamma=delta=0.2
    Hand-set in Appendix D to emphasize compactness; no sensitivity analysis.
  • Token budgets = 3,000 tokens (SkillsBench), 1,200 tokens (ALFWorld)
    Default selection budgets; budget sensitivity is shown in Figure 4 but the defaults themselves are arbitrary.
  • Retrieval thresholds = recall cap 12, cosine min 0.45, 0.90 of max score
    Hand-set in Appendix D; applied uniformly to retrieval baselines.
  • Minimum macro support threshold m = not reported
    Used in Algorithm 3 and ReZip promotion condition (Eq. 9); exact value absent, so replication requires guessing.
  • Risk threshold eta in ReZip = not reported
    Determines when macros are demoted or split; no value or sensitivity analysis given.
  • RRF smoothing constant k and candidate budget K = not reported
    Used in Eq. 6 and Algorithm 3; not specified in the text.
assumptions (4)
  • domain assumption LLM-based section segmentation and role/contract inference (Sec2Graph) recovers all execution-relevant sections, dependencies, and verifier hooks.
    Section 4.1 converts packages via SegmentSkill and InferRole; the entire compression pipeline trusts these extractions. Appendix B.1 shows extraction is imperfect, so this is load-bearing.
  • domain assumption Preserving recorded boundary signatures, dependency closure, and verifier reachability is sufficient for executable reuse.
    MotifZip accepts a macro only when these checks pass (Algorithm 3); the paper explicitly limits the guarantee to recorded structure in Section 4.2, so unrecorded behavior is assumed irrelevant.
  • domain assumption PathHydrate's task anchoring (AnalyzeTask) faithfully converts a query into goal, outputs, capabilities, inputs, and subgoals.
    Section 4.3 and Appendix F define the prompt; incorrect anchors bias section seeding and contribute to the 26% retrieval failure share in Appendix B.8.
  • standard math Standard graph-grammar rewriting and MDL scoring retain their usual formal properties.
    Eqs. 1-3 and Proposition 1 rely on standard graph grammar and description-length concepts from cited graph mining and compression literature.
invented entities (2)
  • Ported macro nodes
    purpose: Compress recurring contract-valid section motifs into reversible procedural units with typed ports, contract record, and source expansion pointers.
    Internal algorithmic construct; its claimed reversibility is evidenced only by in-paper structural checks and benchmarks, with no external falsifiable signature.
  • Canonical prototypes
    purpose: Link contract-compatible section occurrences across skills without erasing occurrence identity or source pointers.
    Auxiliary graph nodes used for reuse grouping; not independently observable outside the framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries." pith.science (2026). https://pith.science/paper/5NM4GHDI

@misc{pith2026260805604,
  author       = {Pith},
  title        = {Pith review of: SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NM4GHDI}},
  note         = {Machine review of arXiv:2608.05604}
}
read the original abstract

Large Language Models (LLMs) increasingly act as agents whose procedural knowledge is stored in reusable skill packages and loaded at inference time. As skill libraries grow, a central challenge is to expose the smallest sufficient executable context under a limited context budget. Existing systems struggle to reuse routines below the whole-skill level, preserve procedural contracts during compression, keep compressed routines executable and expandable, and update the compressed library as skills evolve. These challenges reveal a unit mismatch: skills are retrieved as packages, compressed as text, and converted into execution graphs only after retrieval, whereas reliable reuse requires a contract-bearing procedural unit. We propose SkillZip, an execution-aware procedural abstraction framework that performs contract-preserving compression over section-level graphs. SkillZip rewrites recurring contract-valid motifs into reversible ported macros while preserving boundary signatures, dependency closure, verifier reachability, and source-level expansion. At inference time, it hydrates a compact, dependency-closed context and expands macros only when required. ReZip further integrates new skills and revises risky macros using execution evidence. Comprehensive experiments1 on technical and embodied agent benchmarks show SkillZip consistently outperforms the strongest baseline by up to 12.2 points, while achieving a 3.46x compression ratio with 99.2% dependency preservation and 98.7% verifier reachability. Scaling analyses further confirm robust retrieval across skill libraries ranging from 200 to 100K skills.

Figures

Figures reproduced from arXiv: 2608.05604 by the authors.

Figure 1
Figure 1. Representative skill-library workflows. time [51], allowing procedures to change without retraining. At in￾ference time, a skill provider selects task-relevant procedures from the library and exposes them to the agent. As the library grows, the provider must distinguish among more overlapping packages while avoiding incomplete or redundant context. This raises a central question. How can a skill provider retrieve an… view at source ↗
Figure 2
Figure 2. Overview of the SkillZip framework. Sec2Graph retains occurrence-specific sections and links compatible ones through [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Contract-extraction quality on the annotated subset. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Task reward of SkillZip as the procedural-content [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Mean rendered context on the 1K-skill Skills￾Bench library. Whole-skill comparisons render each retrieved package in full, whereas SkillZip hydrates dependency-closed sections. (RQ11) Is hydrated context compact under the default bud￾get? To measure context compactness…
Figure 7
Figure 7. Figure 7: Failure attribution grouped by the earliest blocking [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 24 canonical work pages

  1. [1]

    Tong Bai, Zhenglin Wan, Pengfei Zhou, Xingrui Yu, Yang You, and Ivor W. Tsang

    work page

  2. [2]

    Julia Belikova, Rauf Parchiev, Evgeny Egorov, Grigorii Davydenko, Gleb Gu- sev, Andrey Savchenko, and Maksim Makarenko. 2026. Managing Procedural Memory in LLM Agents: Control, Adaptation, and Evaluation.arXiv preprint arXiv:2606.23127(2026)

    work page Pith review arXiv 2026

  3. [3]

    Zachary Blumenfeld and Jim Webber. 2026. AIP: A Graph Representation for Learning and Governing Agent Skills.arXiv preprint arXiv:2606.04781(2026)

    work page Pith review arXiv 2026

  4. [4]

    Hongcheol Cho, Ryangkyung Kang, and Youngeun Kim. 2026. SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents.arXiv preprint arXiv:2605.05726(2026). https://arxiv.org/abs/2605.05726

    arXiv 2026

  5. [5]

    Cook and Lawrence B

    Diane J. Cook and Lawrence B. Holder. 1994. Substructure Discovery Using Minimum Description Length and Background Knowledge.Journal of Artificial Intelligence Research1 (1994), 231–255. doi:10.1613/jair.43

    work page doi:10.1613/jair.43 1994

  6. [6]

    Marc-Alexandre Côté, Ákos Kádár, Xingdi Yuan, Ben Kybartas, Tavian Barnes, Emery Fine, James Moore, Matthew Hausknecht, Layla El Asri, Mahmoud Adada, et al. 2018. TextWorld: A Learning Environment for Text-Based Games.arXiv preprint arXiv:1806.11532(2018). https://arxiv.org/abs/1806.11532

    arXiv 2018

  7. [7]

    Tao Feng, Haozhen Zhang, Zijie Lei, Peixuan Han, and Jiaxuan You. 2026. GraphPlanner: Graph Memory-Augmented Agentic Routing for Multi-Agent LLMs.International Conference on Learning Representations(2026). https: //openreview.net/forum?id=ZdGB7MNQDT

    work page 2026

  8. [8]

    Yudong Gao, Zongjie Li, Yuanyuan Yuan, Zimo Ji, Pingchuan Ma, and Shuai Wang. 2026. SkillReducer: Optimizing LLM Agent Skills for Token Efficiency. arXiv preprint arXiv:2603.29919(2026). https://arxiv.org/abs/2603.29919

    arXiv 2026

Show all 63 references
  1. [9]

    Tingxu Han, Yi Zhang, Wei Song, Chunrong Fang, Zhenyu Chen, Youcheng Sun, and Lijie Hu. 2026. SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering?arXiv preprint arXiv:2603.15401(2026)

    2026

  2. [10]

    Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2023. Llmlingua: Compressing prompts for accelerated inference of large language models. InProceedings of the 2023 conference on empirical methods in natural language processing. 13358–13376

    2023

  3. [11]

    Huiqiang Jiang, Qianhui Wu, Xufang Luo, Dongsheng Li, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. 2024. Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression. InProceedings of the 62nd Annual Meeting of the Association for Computational L...

    2024

  4. [13]

    Danai Koutra, U Kang, Jilles Vreeken, and Christos Faloutsos. 2014. VoG: Sum- marizing and Understanding Large Graphs. InProceedings of the 2014 SIAM International Conference on Data Mining. 91–99. doi:10.1137/1.9781611973440.11

    2014 doi

  5. [14]

    Kyoungmin Lee, Hyeonji Jo, Jihoon Ko, Sungsu Lim, and Kijung Shin. 2020. SSumM: Sparse Summarization of Massive Graphs. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 144–154. doi:10.1145/3394486.3403057

    2020

  6. [15]

    Kyoungmin Lee, Jihoon Ko, and Kijung Shin. 2022. SLUGGER: Lossless Hierarchi- cal Summarization of Massive Graphs. In2022 IEEE 38th International Conference on Data Engineering. 2878–2890. https://arxiv.org/abs/2112.05374

    2022 arXiv

  7. [16]

    Kristen LeFevre and Evimaria Terzi. 2010. GraSS: Graph Structure Summarization. InProceedings of the 2010 SIAM International Conference on Data Mining. 454–465. doi:10.1137/1.9781611972801.40

    2010 doi

  8. [17]

    Hao Li, Chunjiang Mu, Jianhao Chen, Siyue Ren, Zhiyao Cui, Yiqun Zhang, Lei Bai, and Shuyue Hu. 2026. Organizing, Orchestrating, and Benchmarking Agent Skills at Ecosystem Scale.arXiv preprint arXiv:2603.02176(2026). https: //arxiv.org/abs/2603.02176

    2026

  9. [18]

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. 2023. Api-bank: A comprehensive benchmark for tool-augmented llms. InProceedings of the 2023 conference on empirical methods in natural language processing. 3102–3116

    2023

  10. [19]

    Xiaoyuan Li, Moxin Li, Keqin Bao, Yubo Ma, Wenjie Wang, Dayiheng Liu, and Fuli Feng. 2026. SkillGraph: Skill-Augmented Reinforcement Learning for Agents via Evolving Skill Graphs.arXiv preprint arXiv:2605.12039(2026). https://arxiv. org/abs/2605.12039

    2026 arXiv

  11. [20]

    Xiangyi Li, Yimin Liu, Wenbo Chen, Bingran You, Zonglin Di, Yifeng He, Shenghan Zheng, et al . 2026. SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks.arXiv preprint arXiv:2602.12670(2026). https://arxiv.org/abs/2602.12670

    2026 arXiv

  12. [21]

    Qiliang Liang, Hansi Wang, Zhong Liang, and Yang Liu. 2026. From skill text to skill structure: The scheduling-structural-logical representation for agent skills. arXiv preprint arXiv:2604.24026(2026)

    2026 arXiv

  13. [22]

    Yuan Liang, Ruobin Zhong, Haoming Xu, Chen Jiang, Yi Zhong, Runnan Fang, Jia-Chen Gu, Shumin Deng, et al. 2026. SkillNet: Create, Evaluate, and Connect AI Skills.arXiv preprint arXiv:2603.04448(2026). https://arxiv.org/abs/2603.04448

    2026

  14. [23]

    George Ling, Shanshan Zhong, and Richard Huang. 2026. Agent Skills: A Data- Driven Analysis of Claude Skills for Extending Large Language Model Function- ality.arXiv preprint arXiv:2602.08004(2026). https://arxiv.org/abs/2602.08004

    2026

  15. [24]

    Dawei Liu, Zongxia Li, Hongyang Du, Xiyang Wu, Shihang Gui, Yongbei Kuang, and Lichao Sun. 2026. Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills.arXiv preprint arXiv:2604.05333(2026). https://arxiv. org/abs/2604.05333

    2026 arXiv

  16. [25]

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. 2024. Agentbench: Evaluating llms as agents. InInternational Conference on Learning Representations, Vol. 2024. 52989–53046

    2024

  17. [26]

    Sebastian Maneth and Fabian Peternek. 2018. Grammar-based graph compression. Information Systems76 (2018), 19–45. doi:10.1016/j.is.2018.03.002

    2018 doi

  18. [27]

    Xiangcheng Meng, Shu Wang, and Yixiang Fang. 2026. SkillRAE: Agent Skill- Based Context Compilation for Retrieval-Augmented Execution.arXiv preprint arXiv:2605.10114(2026). https://arxiv.org/abs/2605.10114

    2026 arXiv

  19. [28]

    Qirui Mi et al . 2026. Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents.arXiv preprint arXiv:2602.01869(2026). https://arxiv.org/abs/2602.01869 Accepted at ICML 2026

    2026 arXiv

  20. [29]

    Yongliang Miao, Ziyang Yu, Liang Zhao, Bowen Zhu, and Hasibul Haque. 2026. SkillLens: Adaptive Multi-Granularity Skill Reuse for Cost-Efficient LLM Agents. arXiv preprint arXiv:2605.08386(2026). https://arxiv.org/abs/2605.08386

    2026 arXiv

  21. [30]

    Siegfried Nijssen and Joost N. Kok. 2004. A Quickstart in Frequent Struc- ture Mining Can Make a Difference. InProceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 647–652. doi:10.1145/1014052.1014134

    2004

  22. [31]

    Siru Ouyang et al . 2026. ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory.International Conference on Learning Representations(2026). https://openreview.net/forum?id=jL7fwchScm

    2026

  23. [32]

    Zhuoshi Pan et al. 2024. LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compression.arXiv preprint arXiv:2403.12968(2024). https://arxiv.org/abs/2403.12968

    2024 arXiv

  24. [33]

    Hongji Pu, Xinyuan Song, and Liang Zhao. 2026. SkillOps: Managing LLM Agent Skill Libraries as Self-Maintaining Software Ecosystems.arXiv preprint arXiv:2605.13716(2026). https://arxiv.org/abs/2605.13716

    2026 arXiv

  25. [34]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun

  26. [35]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: 9 Xingyu Tan et al. Language Models Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Systems. htt...

    2023 arXiv

  27. [36]

    Shuaike Shen, Wenduo Cheng, Mingqian Ma, Alistair Turcan, Martin Jinye Zhang, and Jian Ma. 2026. SKILLFOUNDRY: Building Self-Evolving Agent Skill Libraries from Heterogeneous Scientific Resources.arXiv preprint arXiv:2604.03964(2026). https://arxiv.org/abs/2604.03964

    2026 arXiv

  28. [37]

    Kijung Shin, Amol Ghoting, Myunghwan Kim, and Hema Raghavan. 2019. SWeG: Lossless and Lossy Summarization of Web-Scale Graphs. InProceedings of The Web Conference. 1679–1690. doi:10.1145/3308558.3313402

    2019

  29. [38]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. InAdvances in Neural Information Processing Systems. https://arxiv. org/abs/2303.11366

    2023 arXiv

  30. [39]

    Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. 2020. ALFRED: A Bench- mark for Interpreting Grounded Instructions for Everyday Tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and P...

    2020

  31. [40]

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Cote, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. 2021. ALFWorld: Aligning Text and Em- bodied Environments for Interactive Learning. InInternational Conference on Learning Representations. https://openreview.net/forum?id=0...

    2021

  32. [41]

    Weihang Su, Jianming Long, Qingyao Ai, Yichen Tang, Changyue Wang, Yiteng Tu, and Yiqun Liu. 2026. Skill Retrieval Augmentation for Agentic AI.arXiv preprint arXiv:2604.24594(2026). https://arxiv.org/abs/2604.24594

    2026 arXiv

  33. [42]

    Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, and Wenjie Zhang

  34. [43]

    Hankins, and Jignesh M

    Yuanyuan Tian, Richard A. Hankins, and Jignesh M. Patel. 2008. Efficient Aggre- gation for Graph Summarization. InProceedings of the 2008 ACM SIGMOD Interna- tional Conference on Management of Data. 567–580. doi:10.1145/1376616.1376675

    2008

  35. [44]

    Changyue Wang, Weihang Su, Qingyao Ai, Yichen Tang, Runzhong Qiao, Xu- ancheng Li, Min Zhang, and Yiqun Liu. 2026. Adaptive Multi-Resolution Pro- cedural Knowledge Compression for Large Language Models.arXiv preprint arXiv:2606.12203(2026). https://arxiv.org/abs/2606.12203

    2026 arXiv

  36. [45]

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An Open-Ended Embodied Agent with Large Language Models.arXiv preprint arXiv:2305.16291(2023). https://arxiv.org/abs/2305.16291

    2023 arXiv

  37. [46]

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. 2024. Agent Workflow Memory.arXiv preprint arXiv:2409.07429(2024). https://arxiv.org/ abs/2409.07429

    2024 arXiv

  38. [47]

    Peilin Wu, Mian Zhang, Kun Wan, Wentian Zhao, Kaiyu He, Xinya Du, and Zhiyu Chen. 2026. HiPRAG: Hierarchical Process Rewards for Efficient Agentic Retrieval Augmented Generation.International Conference on Learning Represen- tations(2026). https://openreview.net/forum?id=Gt4v9WBPzm

    2026

  39. [48]

    Tianle Xia, Lingxiang Hu, Yiding Sun, Ming Xu, Lan Xu, Siying Wang, Wei Xu, and Jie Jiang. 2026. GraSP: Graph-Structured Skill Compositions for LLM Agents. arXiv preprint arXiv:2604.17870(2026). https://arxiv.org/abs/2604.17870

    2026 arXiv

  40. [49]

    Zhishang Xiang et al. 2026. When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation.International Conference on Learning Representations(2026). https://openreview.net/forum?id=i9q9xDMjG7

    2026

  41. [50]

    Qinghua Xing, Yinda Chen, Yaping Jin, Zhenhe Wu, Bohan Lin, Hang Zhou, Xinghao Chen, Hanting Chen, and Zhiwei Xiong. 2026. What Should a Skill Remember? Quality–Cost Trade-offs in Cost-Aware Skill Rewriting for Language Model Agents.arXiv preprint arXiv:2606.09421(2026)

    2026 arXiv

  42. [51]

    Renjun Xu and Yang Yan. 2026. Agent skills for large language models: Architec- ture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430 (2026)

    2026 arXiv

  43. [52]

    Xifeng Yan and Jiawei Han. 2002. gSpan: Graph-Based Substructure Pattern Mining. InProceedings of the 2002 IEEE International Conference on Data Mining. 721–724. doi:10.1109/ICDM.2002.1184038

    2002 arXiv

  44. [53]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. In NeurIPS 2022 Foundation Models for Decision Making Workshop

    2022

  45. [54]

    Kun Zeng et al. 2026. Group of Skills: Group-Structured Skill Retrieval for Agent Skill Libraries.arXiv preprint arXiv:2605.06978(2026). https://arxiv.org/abs/2605. 06978

    2026 arXiv

  46. [55]

    Guibin Zhang, Muxin Fu, and Shuicheng Yan. 2026. MemGen: Weaving Gen- erative Latent Memory for Self-Evolving Agents.International Conference on Learning Representations(2026). https://openreview.net/forum?id=vI56m4Iu4e

    2026

  47. [56]

    Xing Zhang, Guanghui Wang, Yanwei Cui, Wei Qiu, Ziyuan Li, Bing Zhu, and Peiyang He. 2026. Experience compression spectrum: Unifying memory, skills, and rules in LLM agents.arXiv preprint arXiv:2604.15877(2026)

    2026 arXiv

  48. [57]

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. ExpeL: LLM Agents Are Experiential Learners. InProceedings of the AAAI Conference on Artificial Intelligence. https://arxiv.org/abs/2308.10144

    2024 arXiv

  49. [58]

    Fatemi, Xiaolong Jin, Zora Zhiruo Wang, Apurva Gandhi, Yueqi Song, Yu Gu, Jayanth Srinivasa, Gaowen Liu, Graham Neubig, and Yu Su

    Boyuan Zheng, Michael Y. Fatemi, Xiaolong Jin, Zora Zhiruo Wang, Apurva Gandhi, Yueqi Song, Yu Gu, Jayanth Srinivasa, Gaowen Liu, Graham Neubig, and Yu Su. 2025. SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills.arXiv preprint arXiv:2504.07079(2025). ht...

    2025 arXiv

  50. [59]

    Yingli Zhou. 2026. A comprehensive survey on agent skills: Taxonomy, tech- niques, and applications.Techniques, and Applications (April 27, 2026)(2026)

    2026

  51. [60]

    Yifan Zhou et al. 2026. Benchmarking Skill Generation Pipelines for LLM Agents. arXiv preprint arXiv:2605.18693(2026). https://arxiv.org/abs/2605.18693

    2026 arXiv

  52. [61]

    normalize the headers in a CSV file and verify that the row count is unchanged

    Zijian Zhou et al . 2026. MEM1: Learning to Synergize Memory and Reason- ing for Efficient Long-Horizon Agents.International Conference on Learning Representations(2026). https://openreview.net/forum?id=XY8AaxDSLb 10 SkillZip: Contract-Preserving Graph Compression for Scalable...

    2026

  53. [2024]

    InInternational Conference on Learning Representations

    ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs. InInternational Conference on Learning Representations

  54. [2025]

    InProceedings of the ACM on Web Conference 2025

    Paths-over-graph: Knowledge graph empowered large language model reasoning. InProceedings of the ACM on Web Conference 2025. 3505–3522

    2025

  55. [2026]

    https://arxiv.org/abs/2606.03056

    SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale.arXiv preprint arXiv:2606.03056(2026). https://arxiv.org/abs/2606.03056

    2026 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.