Pith. sign in

REVIEW 3 major objections 5 minor 23 references

ArtisanCAD uses CAD-IR to turn expert CAD procedures into reusable skills and to convert vague design requests into editable industrial B-Rep models.

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

2026-07-11 02:33 UTC pith:5GSMZEXN

load-bearing objection Solid industrial CAD-agent systems paper: CAD-IR scaffold is measured; expert skill transfer is real but only shown qualitatively on four parts. the 3 major comments →

arxiv 2607.05750 v2 pith:5GSMZEXN submitted 2026-07-07 cs.AI cs.GR

ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation

Yunhan Xu , Qifeng Wu , Xunjin Li , Yuanwei Bin , Qingsong Yao , Jianghang Gu , Guan Wang , Weihao Lv
show 5 more authors
Huiyu Yang Wenfa Luo Jiao Xiang Yuntian Chen Shiyi Chen
This is my paper
classification cs.AI cs.GR
keywords CAD agentcomputer-aided designindustrial designgeometric modelingCAD intermediate representationexpert knowledge distillationtext-to-CADB-Rep
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.

Industrial CAD is not one-shot shape generation: it needs long ordered procedures, feature dependencies, parameters that stay editable, and production B-Rep execution. Existing text-to-CAD systems struggle when prompts are ambiguous or only state high-level intent, and they rarely reuse the procedural knowledge that experts already leave in macro logs, drawings, and operation histories. This paper presents ArtisanCAD, a skill-guided CAD agent built around CAD-IR, an executable intermediate representation that records parameters, ordered operations, tool bindings, dependencies, entities, and checks. CAD-IR both distills expert CATIA recordings into parameterized skills and scaffolds incomplete prompts into complete runnable procedures, which a CATIA backend then executes and a multi-view vision loop can revise. On public intermediate prompts, CAD-IR alone lowers mean Chamfer Distance from 14.83 to 9.88; on four automotive part families, skill-guided CAD-IR yields editable CATIA-native variants where generation without those skills fails.

Core claim

The paper claims that an explicit executable intermediate representation, CAD-IR, is the missing procedural layer for industrial CAD agents: it carries expert macro-derived skills and also structures vague or intermediate text into complete CAD operations. With CAD-IR, intermediate-prompt generation on Text2CAD improves (mean CD 14.83 to 9.88; solid IoU 0.614 to 0.646 without expert skills), and expert CATIA recordings distilled into skills enable new variant requests for complex automotive components to produce editable production-oriented B-Rep models via CATIA-MCP—something the agent cannot reliably do from short requests alone.

What carries the argument

CAD-IR (CAD intermediate representation): an executable procedural scaffold Z = (P, T, O, E, V) of parameters, backend tools, ordered operations, entities/dependencies, and verification rules. Each operation binds intent, MCP calls, arguments, outputs, dependencies, and checks. It is the carrier for expert-to-skill distillation and the object the agent instantiates, rewrites, executes, and visually revises.

Load-bearing premise

The approach assumes that a few expert-derived part skills plus multi-view visual PASS/FAIL rewriting are enough for real production variants—that expert operation order and template IR stay valid scaffolds and that visual feedback catches structural failures without broader industrial metrics or many more part families.

What would settle it

Take a held-out industrial component family with no skill template, or variants that break the expert operation order, and check whether ArtisanCAD still completes editable CATIA-native B-Reps that match expert geometry and feature trees; if it systematically fails while skill-guided cases succeed only for the four trained families, the central generalization claim fails.

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

If this is right

  • Ambiguous intermediate text can be turned into executable CAD without first writing expert-level step-by-step prompts.
  • CATIA macros, drawing notes, and parameter tables become reusable parameterized skills for new variants in the same part family.
  • Planning, execution, and repair can stay at the IR level while a CATIA-MCP (or other) backend produces editable production B-Reps.
  • Without skill-guided CAD-IR, long-horizon industrial parts remain out of reach for short variant requests alone.

Where Pith is reading between the lines

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

  • If CAD-IR is the real bottleneck, the same skill-plus-IR pattern should transfer to other industrial CAD systems once tool bindings exist, not only CATIA.
  • Vision-only PASS/FAIL may miss tolerance, manufacturability, and assembly-interface failures; adding those checks would be a direct next stress test.
  • A larger multi-family skill library could turn one-off expert demos into a shared industrial procedure catalog rather than four isolated templates.

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

3 major / 5 minor

Summary. ArtisanCAD is a skill-guided industrial CAD agent whose core is CAD-IR, an executable intermediate representation encoding parameters, ordered operations, MCP tool bindings, dependencies, generated entities, and verification rules. CAD-IR is used offline to distill expert CATIA macros, drawing notes, and engineering descriptions into reusable part skills s=(Ds,Os,Ps,Zs), and online as a procedural scaffold that turns ambiguous or intermediate prompts into complete CAD procedures. The agent retrieves a skill, instantiates/revises CAD-IR (via Mimo-v2.5-Pro), executes through a CATIA-MCP backend, and refines with multi-view visual feedback. On 100 Text2CAD intermediate prompts (no expert skills), CAD-IR reduces mean CD from 14.83 to 9.88 and raises solid IoU from 0.614 to 0.646 versus a no-IR ablation, with further gains over Text2CAD and CAD-Coder (Table 2, Fig. 3). On four automotive component families, expert macros are distilled into skills and used to generate editable CATIA-native B-Rep variants for expert-provided requests (Fig. 4); the paper states that generation without skill-guided CAD-IR fails on these long-horizon parts.

Significance. If the dual claim holds, the paper offers a concrete path from public text-to-CAD toward production CAD automation: an explicit executable IR that both improves intermediate-prompt generation without expert knowledge and carries expert procedural knowledge into reusable skills for industrial B-Rep variants. Strengths include a clear IR design (Eqs. 1–3), a dedicated industrial CATIA-MCP backend producing editable B-Rep, a controlled public-benchmark ablation isolating CAD-IR (Table 2), and qualitative industrial demos on real automotive parts. The work is timely for CAD agents and industrial AI, and the multi-backend IR design is a useful systems contribution even if industrial evaluation remains limited.

major comments (3)
  1. Sec. 4.2 / Fig. 4: The industrial half of the central claim—that expert-to-skill distillation via CAD-IR enables production-oriented variant generation where skill-free generation fails—is only qualitative. The paper asserts failure without skill-guided CAD-IR but reports no success/failure rates, no CD/IoU or feature-tree/editability metrics, no held-out variant set, and no quantitative ablation of skill retrieval vs. IR-only vs. vision rewrite. For a load-bearing claim about industrial transfer, at least a small controlled table (e.g., success rate and one geometric or structural metric over the reported variants, with and without skills) is needed.
  2. Table 2: Public-benchmark evidence for CAD-IR is the strongest result but is under-specified for robustness. Evaluation uses 100 randomly sampled intermediate prompts with no error bars, confidence intervals, or multiple seeds; CD and solid IoU are reported without variance. Given that the abstract and contributions lead with the 14.83→9.88 CD improvement, either expand the sample / report uncertainty or clearly frame the result as a fixed-split demonstration rather than a definitive ranking.
  3. Sec. 3.3.3 and Sec. 4.2: Multi-view visual PASS/FAIL is the primary industrial refinement signal, yet the paper does not validate that eight-view boards reliably detect structural or feature-dependency failures (as opposed to coarse shape errors). Without even a small agreement study or failure-mode analysis against expert checks, the claim that vision-guided IR rewrite yields production-ready models remains weakly supported.
minor comments (5)
  1. Table 1 and related work: Several baselines (CADFusion, CAD-Assistant, FutureCAD) are listed with venue years including 2026; ensure citation status and arXiv identifiers are accurate and consistent with the reference list.
  2. Eqs. (1)–(3): Notation for CAD-IR and skills is clear, but the manuscript never shows a full example IR JSON or operation schema in the main text; a short appendix listing one complete operation would aid reproducibility.
  3. Fig. 1 and Fig. 2: Overview figures are informative but dense; labels such as “MCP” and “IR rewrite” could be expanded once in captions for readers outside the agent-tooling community.
  4. Sec. 4.1 metrics: Clarify the exact CD scaling (CD×10^3) and solid IoU resolution (32^3) in the table caption as well as the text, and state whether meshes are normalized before comparison.
  5. Acknowledgments and author list: “Acknowledgments will be added in the final version” is fine for arXiv but should be completed for journal submission; also check equal-contribution and corresponding-author formatting consistency.

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with external benchmarks and an IR ablation; industrial demos are skill-reuse evaluation, not results forced by construction.

full rationale

ArtisanCAD is an engineering/systems paper, not a first-principles derivation. Its load-bearing claims are empirical: (1) CAD-IR improves intermediate-prompt generation on Text2CAD (Table 2: mean CD 14.83→9.88, solid IoU 0.614→0.646 vs the same agent without IR; also vs Text2CAD and CAD-Coder), and (2) expert CATIA macros can be distilled into parameterized skills that support variant generation on four automotive families (Sec. 4.2, Fig. 4). Claim (1) is measured against external baselines and an internal ablation without expert skills, so it is not defined into existence by the inputs. Claim (2) reuses expert-derived templates within the same component family by design of skill transfer; success is not mathematically forced (the paper asserts failure without skill-guided IR), though evaluation is qualitative. There is no self-definitional equation, no fitted parameter renamed as a prediction, no load-bearing uniqueness theorem from overlapping authors, and no ansatz smuggled in via self-citation. Related skill papers [21–22] are not author-overlapping uniqueness claims that force the result. Weak industrial metrics and same-family skill reuse are evaluation limitations, not circularity under the stated criteria.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central claims rest on engineering design choices and empirical evaluation assumptions rather than free physical constants. Load-bearing premises are that expert macros can be normalized into reusable parameterized skills, that CAD-IR is a sufficient executable scaffold for both public and industrial settings, that CATIA-MCP faithfully realizes IR operations as editable B-Rep, and that multi-view LLM visual review is an adequate repair signal. Free parameters are mostly system knobs (sample size, views, models, refinement budget) rather than fitted scientific constants.

free parameters (4)
  • Text2CAD evaluation sample size (100 intermediate prompts)
    Central public-benchmark claim depends on this hand-chosen random subset size; no full-split or variance reporting.
  • Multi-view visual board composition (8 canonical views)
    Feedback loop uses a fixed eight-view set; choice affects what failures the rewrite model can see.
  • Maximum IR refinement budget
    Iterative rewrite stops at a budget; value is not quantified but affects success/failure of generation.
  • LLM backends (Codex-xHigh for skill acquisition; Mimo-v2.5-Pro for IR rewrite/vision)
    Results depend on specific proprietary model choices and prompts not fully specified as fixed artifacts.
axioms (5)
  • domain assumption Expert CATIA macro order and parameter schemas are transferable priors for new variants within a component family.
    Core of expert-to-skill acquisition (Sec. 3.2); if macros are brittle or non-transferable, skill-guided generation fails.
  • ad hoc to paper An intermediate executable IR (parameters, ordered MCP calls, dependencies, entities, verification) is a better planning target than direct backend scripts or raw command sequences.
    Central design choice of CAD-IR (Sec. 3.1–3.2); supported by ablation but not independently proven as a general principle.
  • domain assumption Multi-view renderings plus an LLM visual judge can detect geometric/structural intent failures sufficiently to guide IR repair.
    Vision feedback loop (Sec. 3.3.3) is load-bearing for iterative refinement quality.
  • domain assumption Chamfer Distance (×10^3) and solid IoU at 32^3 are adequate proxies for CAD generation quality on Text2CAD.
    Standard in cited CAD generation evaluation; used as primary public evidence (Sec. 4.1).
  • domain assumption CATIA-MCP tool bindings faithfully implement CAD-IR operations as editable production-oriented B-Rep features.
    Backend execution assumption (Sec. 3.3.2); industrial claims depend on this interface correctness.
invented entities (3)
  • CAD-IR (CAD intermediate representation) no independent evidence
    purpose: Unified executable procedural scaffold encoding parameters, tools, ordered operations, entities/dependencies, and verification for distillation and generation.
    Primary invented representation; independent evidence is empirical (benchmark ablation + industrial demos), not external formal validation.
  • Part skill s = (Ds, Os, Ps, Zs) no independent evidence
    purpose: Reusable package of description, expert operation order, parameter schema, and template IR distilled from expert records.
    Organizes expert knowledge transfer; evidence is internal case studies on four components.
  • CATIA-MCP backend no independent evidence
    purpose: Maps CAD-IR operations to CATIA-native MCP tool calls for industrial B-Rep execution.
    Enabling system component rather than a scientific entity; not independently released/validated outside the paper.

pith-pipeline@v1.1.0-grok45 · 17241 in / 3628 out tokens · 43364 ms · 2026-07-11T02:33:17.292830+00:00 · methodology

0 comments
read the original abstract

Computer-aided design (CAD) for industrial components requires long-horizon procedural modeling, robust feature dependencies, editable parametric geometry, and production-grade B-Rep execution. Existing text-to-CAD methods have made promising progress in generating CAD programs from natural-language descriptions, but they still struggle when user prompts are ambiguous, underspecified, or only describe high-level design intent. They also rarely exploit expert procedural knowledge naturally available in industrial workflows, such as CATIA operation recordings, macro logs, drawing notes, and engineering descriptions. We present ArtisanCAD, a skill-guided industrial CAD agent with expert-grounded knowledge distillation. The core of ArtisanCAD is CAD intermediate representation (CAD-IR), an executable procedural representation that encodes parameters, ordered operations, MCP tool bindings, dependencies, generated entities, and verification rules. CAD-IR plays two key roles: it first serves as the carrier for distilling expert CAD procedures into reusable parameterized skills; then it provides a procedural scaffold that turns vague or intermediate-level prompts into complete executable CAD operations. ArtisanCAD retrieves expert-derived skills, instantiates and revises CAD-IR, executes the resulting procedure through a dedicated CATIA-MCP backend, and uses multi-view visual feedback for iterative refinement, and finally generates production-ready B-Rep models. On the Text2CAD benchmark, CAD-IR improves generation from intermediate prompts by reducing mean Chamfer Distance from $14.83$ to $9.88$, showing its ability to bridge ambiguous textual intent and executable CAD construction. On four complex automotive components, CAD-IR enables expert CATIA recordings to be distilled into reusable skills, allowing ArtisanCAD to generate editable CATIA-native B-Rep models for new variant requests.

Figures

Figures reproduced from arXiv: 2607.05750 by Guan Wang, Huiyu Yang, Jianghang Gu, Jiao Xiang, Qifeng Wu, Qingsong Yao, Shiyi Chen, Weihao Lv, Wenfa Luo, Xunjin Li, Yuanwei Bin, Yunhan Xu, Yuntian Chen.

Figure 1
Figure 1. Figure 1: Overview of ARTISANCAD . Expert CATIA recordings and drawing notes are distilled into reusable part skills with parameters, checks, template CAD-IR, and ordered operations. Given ambiguous variant requests, ARTISANCAD retrieves the corresponding skill, rewrites CAD-IR through an agent loop, executes it via CAD-MCP, and produces evaluated, editable, production-ready B-Rep CAD models. tolerances, wall thickn… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the skill-guided variant CAD generation framework. Given an ambiguous or variant user [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on representative Text2CAD samples. Each row shows one target shape, and each [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Industrial variant generation with expert-derived skills. Each row corresponds to one automotive component [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

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

23 extracted references · 23 canonical work pages · 7 internal anchors

  1. [1]

    John Wiley & Sons, 1995

    Jami J Shah and Martti Mäntylä.Parametric and feature-based CAD/CAM: concepts, techniques, and applications. John Wiley & Sons, 1995

    work page 1995

  2. [2]

    Parametric cad modeling: An analysis of strategies for design reusability.Computer-aided design, 74:18–31, 2016

    Jorge D Camba, Manuel Contero, and Pedro Company. Parametric cad modeling: An analysis of strategies for design reusability.Computer-aided design, 74:18–31, 2016

    work page 2016

  3. [3]

    SketchGraphs: A Large-Scale Dataset for Modeling Relational Geometry in Computer-Aided Design

    Ari Seff, Yaniv Ovadia, Wenda Zhou, and Ryan P Adams. Sketchgraphs: A large-scale dataset for modeling relational geometry in computer-aided design.arXiv preprint arXiv:2007.08506, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  4. [4]

    Computer-aided design as language

    Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. Advances in Neural Information Processing Systems, 34:5885–5897, 2021

    work page 2021

  5. [5]

    Sketchgen: Generating constrained cad sketches.Advances in Neural Information Processing Systems, 34:5077–5088, 2021

    Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches.Advances in Neural Information Processing Systems, 34:5077–5088, 2021

    work page 2021

  6. [6]

    Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences.ACM Transactions on Graphics (TOG), 40(4):1–24, 2021

    Karl DD Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences.ACM Transactions on Graphics (TOG), 40(4):1–24, 2021

    work page 2021

  7. [7]

    Deepcad: A deep generative network for computer-aided design models

    Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. InProceedings of the IEEE/CVF international conference on computer vision, pages 6772–6782, 2021

    work page 2021

  8. [8]

    SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks

    Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks.arXiv preprint arXiv:2207.04632, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  9. [9]

    Free2cad: Parsing freehand drawings into cad commands.ACM Transactions on Graphics (TOG), 41(4):1–16, 2022

    Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J Mitra. Free2cad: Parsing freehand drawings into cad commands.ACM Transactions on Graphics (TOG), 41(4):1–16, 2022

    work page 2022

  10. [10]

    Csgnet: Neural shape parser for constructive solid geometry

    Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. Csgnet: Neural shape parser for constructive solid geometry. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5515–5523, 2018

    work page 2018

  11. [11]

    Shapeassembly: Learning to generate programs for 3d shape structure synthesis.ACM Transactions on Graphics (TOG), 39(6):1–20, 2020

    R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis.ACM Transactions on Graphics (TOG), 39(6):1–20, 2020

    work page 2020

  12. [12]

    Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts.Advances in Neural Information Processing Systems, 37:7552–7579, 2024

    Mohammad S Khan, Sankalp Sinha, Talha U Sheikh, Didier Stricker, Sk A Ali, and Muhammad Z Afzal. Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts.Advances in Neural Information Processing Systems, 37:7552–7579, 2024

    work page 2024

  13. [13]

    Cad translator: An effective drive for text to 3d parametric computer-aided design generative modeling

    Xueyang Li, Yu Song, Yunzhong Lou, and Xiangdong Zhou. Cad translator: An effective drive for text to 3d parametric computer-aided design generative modeling. InProceedings of the 32nd ACM International Conference on Multimedia, pages 8461–8470, 2024. 11 ArtisanCAD

    work page 2024

  14. [14]

    Cad-llama: leveraging large language models for computer-aided design parametric 3d model generation

    Jiahao Li, Weijian Ma, Xueyang Li, Yunzhong Lou, Guichun Zhou, and Xiangdong Zhou. Cad-llama: leveraging large language models for computer-aided design parametric 3d model generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18563–18573, 2025

    work page 2025

  15. [15]

    Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced multimodal llms

    Siyu Wang, Cailian Chen, Xinyi Le, Qimin Xu, Lei Xu, Yanzhou Zhang, and Jie Yang. Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced multimodal llms. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 7880–7888, 2025

    work page 2025

  16. [16]

    Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward.Advances in Neural Information Processing Systems, 38:59765–59789, 2026

    Yandong Guan, Xilin Wang, Ximing Xing, Jing Zhang, Dong Xu, and Qian Yu. Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward.Advances in Neural Information Processing Systems, 38:59765–59789, 2026

    work page 2026

  17. [17]

    Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models

    Ruiyu Wang, Yu Yuan, Shizhao Sun, and Jiang Bian. Text-to-cad generation through infusing visual feedback in large language models.arXiv preprint arXiv:2501.19054, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    Text-to-CadQuery: A New Paradigm for CAD Generation with Scalable Large Model Capabilities

    Haoyang Xie and Feng Ju. Text-to-cadquery: A new paradigm for cad generation with scalable large model capabilities.arXiv preprint arXiv:2505.06507, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Cad-assistant: tool-augmented vllms as generic cad task solvers

    Dimitrios Mallis, Ahmet Serda Karadeniz, Sebastian Cavada, Danila Rukhovich, Niki Foteinopoulou, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-assistant: tool-augmented vllms as generic cad task solvers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7284–7294, 2025

    work page 2025

  20. [20]

    Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding

    Jiahao Li, Qingwang Zhang, Qiuyu Chen, Guozhan Qiu, Yunzhong Lou, and Xiangdong Zhou. Towards high- fidelity cad generation via llm-driven program generation and text-based b-rep primitive grounding.arXiv preprint arXiv:2603.11831, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  21. [21]

    Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills

    Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Erchao Zhao, Xiaoxi Jiang, and Guanjun Jiang. Trace2skill: Distill trajectory-local lessons into transferable agent skills.arXiv preprint arXiv:2603.25158, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [22]

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

    Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, et al. Skillrl: Evolving agents via recursive skill-augmented reinforcement learning.arXiv preprint arXiv:2602.08234, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  23. [23]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017. 12

    work page 2017