Pith. sign in

REVIEW 4 major objections 5 minor 46 references

SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Supervised fine-tuning on SKT's 27,164 verified synthetic trajectories—generated from 4,000 tasks built on 2,000 public skills—improves skill-use performance in all 16 model–benchmark–harness comparisons tested, with gains of 3.20 to 18.91

desk verdict Solid empirical pipeline paper with a consistent training signal, but the 'verified' label leans on an unvalidated LLM judge and SkillEval shares the pipeline's selection criteria. read the letter →

arxiv 2608.02287 v1 pith:SUIPLK42 submitted 2026-08-03 cs.AI

Zelin Tan , Yiqun Zhang , Hao Li , Zhiyao Cui , Hejia Geng , Shao Zhang , Hangfan Zhang , Yang Chen
show 6 more authors
Xiaosong Wang Lilong Wang Zhenfei Yin Shuyue Hu Chen Zhang Lei Bai
This is my paper · ORCID
classification cs.AI
keywords syntheticdatagenerationskill-usetrainingLLMagentssupervisedfine-tuningverificationagentskillsbenchmarkconstruction
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

This paper tries to establish that the missing ingredient in getting language agents to use external skills is not skill content but training data: models can be taught when and how to apply skills through supervised fine-tuning on large amounts of verified, skill-grounded execution traces. The authors build SKT, a three-stage pipeline that turns a pool of 2,000 public skills into 4,000 executable tasks and 27,164 verified trajectories, each filtered by deterministic checks plus LLM judgment to confirm that every required skill was actually used. Across two base models, two agent harnesses, and four benchmarks, SFT on these trajectories improves every one of 16 comparisons, with gains of roughly 3–19 points; the gains shrink when skills are withheld, become negative when verification is removed, and grow with the training-skill budget. If the claim holds, skill-use ability can be trained at scale from synthetic data rather than manual demonstrations, and verification quality—not just trajectory count—is the ingredient that carries the improvement.

What carries the argument

Load-bearing object: the 'verified trajectory' — a complete rollout that passes a deterministic check (full evaluator credit, normal termination, explicit skill access) and an LLM validator confirming each skill was consulted before the action, affected a concrete decision or operation, and was applied correctly. Around it sits the task-package machinery: template-driven synthesis of executable, gradable tasks; rule-based and LLM semantic verification; a paired-rollout skill-dependence test (with-skill score greater than without-skill); and a difficulty gate that sends too-easy tasks back for repair. Verified trajectories become SFT targets with loss masked to assistant reasoning and tool-ca

What would settle it

Take a random sample of the 27,164 accepted trajectories and have human raters apply the paper's trajectory-side criteria (skill consulted before the action, concrete effect, correct application). If a substantial fraction pass the LLM validator but fail human review—or if trajectories that fail human review still produce the same SFT gains—the 'verified' advantage would be explained by verifier leniency rather than genuine skill-use learning.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that verified synthetic data generation is a viable and scalable route to skill-use training: SFT on SKT trajectories improves skill-use performance on every benchmark–model–harness combination tested, and the improvement is attributable to the verification pipeline rather than to raw data volume. The authors demonstrate this by ablating verification (unverified trajectories actually lower scores), by withholding skills at evaluation time (gains mostly disappear), and by scaling the training-skill budget from 100 to 2,000 (scores climb monotonically). They also show the behavior transfers across agent interfaces, retaining about half the ma

Load-bearing premise

The 'verified' label rests on LLM judges—never checked against human labels—who must decide each skill was consulted before the action, affected a concrete decision, and was applied correctly, and that a bundle-level paired-rollout score difference proves skill dependence.

Editorial extensions

If this is right

  • If correct, skill-use ability can be instilled by fine-tuning on machine-generated verified trajectories, without human-written demonstrations.
  • Verification is not a minor cleanup: raw synthetic trajectories actually reduce performance, so the quality-control gates are what produce the gains.
  • Learned skill-use behavior transfers across agent interfaces, retaining about half the matched-harness gain, so a substantial part of the ability is interface-independent.
  • A single checkpoint trained on trajectories from multiple harnesses performs close to harness-specific specialists, reducing the need to maintain one model per execution environment.
  • Performance rises monotonically as the training-skill budget grows from 100 to 2,000 skills, suggesting the pipeline will continue to pay off as more public skills become available.

Reading between the lines

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

  • The 'verified' advantage would be more convincing with a human audit: the LLM validator and task verifier are never calibrated against human labels, so testing verifier leniency would directly probe the claimed mechanism.
  • The same template-driven synthesis-plus-verification pattern could train agents to follow other procedural documents—regulations, runbooks, APIs—not just skill packages.
  • Because gains largely vanish when skills are withheld at inference, SKT teaches use of external instructions rather than memorization of content; a natural next test is whether combining SKT with a model that also retrieves its own skills compounds the gains.
  • SkillEval is generated by the same pipeline as the training data, so its distribution is aligned with the training objective; a benchmark authored by humans outside the pipeline would test the generality claim more strictly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces SKT (Skill-use Training), a three-stage pipeline that curates public agent skills, synthesizes executable skill-grounded tasks with rule-based/agent-based verification and feedback-guided repair, and collects trajectories that pass deterministic and LLM-based skill-use validation. Using 2,000 skills, the pipeline produces 4,000 task packages and 27,164 verified trajectories. Supervised fine-tuning on these trajectories is evaluated on four benchmarks (including the newly introduced SkillEval) across two backbones and two harnesses. The authors report that all 16 matched-harness comparisons improve (Table 1), that gains shrink when skills are withheld (Figure 3), that removing verification eliminates or reverses gains (Figure 4), that cross-harness transfer retains roughly half of the matched gain (Figure 5), and that performance scales with the training-skill budget (Figure 6).

Significance. If the central claims hold, the paper makes a strong, practical contribution: it provides a scalable template for generating verifiable skill-use supervision and shows consistent gains on multiple external benchmarks. The evidence base is unusually broad—16 matched comparisons, two harnesses, two backbones, four benchmarks, cross-harness and scaling analyses—and the appendices give unusually detailed acceptance criteria. The main value would be in establishing that verified synthetic data generation, not merely more synthetic data, improves skill-use. However, that specific claim is currently vulnerable because the 'verified' label depends on an unvalidated LLM judge, and the key ablation (Figure 4) does not isolate the skill-use verification component. SkillEval also shares the synthesis pipeline with the training corpus, introducing a circularity that the external benchmarks only partially offset.

major comments (4)
  1. [§3.3, Appendix B.4] The central 'verified' claim rests on a single LLM validator (DeepSeek V4 Pro) applying the three B.4 criteria: consultation before the guided action, effect on a concrete decision, and correct application. No human-label validation, inter-annotator agreement, or error analysis is reported for this judge. Since every one of the 27,164 trajectories passed (Section 4.1), a systematically lenient judge would make 'verified' vacuous and the Figure 4 advantage attributable to judge bias. I request a human-annotation study on a stratified sample of accepted/rejected trajectories, with agreement metrics and qualitative error analysis, and a statement of which failure modes remain.
  2. [§4.3, Figure 4] The verified-vs-unverified ablation removes verification and repair jointly. The 'Unverified' condition therefore also drops the deterministic success filter (full evaluator credit, normal termination, well-formed trace) and the task-level quality gates. The 11.91–24.61 point gaps could thus reflect training on successful vs failed/partial rollouts rather than the specific skill-use criterion that the paper's title foregrounds. Please add a controlled ablation that applies all deterministic success/trace filters and removes only the LLM skill-use check (or that adds the LLM check to raw trajectories). Also report per-condition variances or tests; Figure 4 shows four-run means without error bars or significance tests.
  3. [§4.1, §4.2, Appendix C.2] SkillEval is generated by the same task-synthesis pipeline—same rubric filter, same rule-based/agent-based verifiers, same difficulty controller—as the SKT training corpus, from a disjoint skill pool. Improvements on SkillEval therefore partly measure agreement with SKT's own selection criteria, which is structural circularity. The largest gains in Table 1 are on SkillEval (10.25–18.91 points), so the headline result depends on this benchmark. The external benchmarks (SkillsBench, MolBench-Bind, AgentSkillOS) mitigate the concern, but an evaluation on an independently constructed skill-use benchmark, or a distribution-shift analysis showing SkillEval is not just an in-distribution artifact, would substantially strengthen the generality claim.
  4. [Table 4, Appendix B.2, §3.2] The skill-dependence gate accepts a task when r+ > r- in paired rollouts at the bundle level. For multi-skill tasks (62% of the pool: k=2 and k=3), this does not establish that every designated skill is necessary or even used; one redundant skill can be carried by the other. The paper also does not report the number of paired rollouts or the stability of r+ > r- across seeds. Since the training tasks are supposed to be 'skill-grounded', this weakens the compositionality claim. Please report the distribution of r+ - r-, the number of rollouts used, and, on a sample, leave-one-skill-out checks to verify that multi-skill tasks indeed depend on each designated skill.
minor comments (5)
  1. [Figure 4] Figure 4 shows only four-run means without error bars or significance tests. Given the SDs in Table 1, the +5.1 gain on AgentSkillOS may be within noise; please add error bars and, where appropriate, paired tests across runs.
  2. [Figure 6] The x-axis uses equal spacing for budgets 0, 100, 500, 1k, 2k, which visually exaggerates early scaling. Use the actual budget values on a log-like axis or mark the axis as categorical.
  3. [Author line] The contact line contains an apparent rendering artifact: 'envel⌢pe' should be 'envelope'.
  4. [Table 1] The 'General-Purpose' column heading is vague; consider 'General' or 'Skill Use' for the four benchmark columns, or add a note that SkillsBench and SkillEval are the main skill-use measures.
  5. [Appendix C.3] The note that 'syntactic counts do not themselves establish successful skill loading, substantive skill use, or trajectory acceptance' is an honest caveat, but it also underlines the need for the validator validation requested in Major Comment 1. Consider moving this caveat to the main text when the trajectory counts are first presented.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: central SFT benefit is supported by external benchmarks and ablations; SkillEval's shared pipeline is a validation-overlap concern, not a by-construction reduction.

full rationale

The paper's derivation chain is empirical rather than definitional: SKT filters trajectories with rule-based and LLM validators (Section 3.3; Appendix B.4) and then measures SFT gains on four benchmarks. No parameter is fitted to the evaluation benchmarks, and no benchmark score is fed back into the pipeline. The three public benchmarks (SkillsBench, MolBench-Bind, AgentSkillOS-Bench) are external to SKT, and the withheld-skills, verification-ablation, cross-harness, and scaling experiments provide independent evidence that the gains track skill availability and data quality. The main overlap is SkillEval, which is explicitly constructed with the same task-synthesis and verification pipeline from a disjoint pool (Section 4.1; Appendix C.2). This means SkillEval shares the SKT selection criteria and could overstate generalization to unrelated skill-use distributions; it is, however, a distributional/validation concern, not a circular reduction of the training objective to the evaluation metric, and the external benchmarks prevent the central claim from resting on SkillEval alone. The unvalidated LLM verifier (Appendix B.4) is a robustness threat, not a circularity. Self-citations (e.g., Li et al. 2026a; Zhang et al. 2026) are to benchmarks and related work, not to a load-bearing uniqueness or ansatz result.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

SKT introduces no new physical or formal entities; it introduces a data-generation pipeline, a dataset, and a benchmark. The central claim rests mainly on hand-chosen pipeline thresholds, the unvalidated reliability of LLM verifiers, and the assumption that external benchmarks faithfully measure skill-use rather than related confounds.

free parameters (4)
  • Difficulty threshold theta_easy = 0.6 (>=3 of 5 full-credit rollouts triggers repair)
    Chosen by hand; controls which tasks are accepted into the training and SkillEval pools, thereby shaping the entire task distribution.
  • Difficulty rollouts N = 5
    Hand-set; the pass-rate estimate p_pass and the accepted-task distribution depend on N.
  • Skill-dependence criterion = r+ > r- for bundle-level paired rollouts
    Any positive score difference counts as skill dependence; no per-skill or thresholded necessity is required. This defines which tasks are admitted and partly defines what 'skill-use' means in the evaluation.
  • SFT recipe (learning rate, epochs, batch) = 5e-6, 1 epoch, batch 8
    Standard recipe, not fitted to benchmarks, but every reported result depends on this training setup.
assumptions (5)
  • domain assumption Agent Skills (SKILL.md packages in skills.sh) are valid, loadable procedural knowledge for target LLM agents.
    The whole pipeline assumes skill files can be meaningfully executed by agents; no independent check of skill-file usability is reported.
  • domain assumption LLM verifiers (DeepSeek V4 Pro in this pipeline) correctly judge task semantics, skill dependence, and faithful skill-use without human validation.
    Sections 3.2 and 3.3 and Appendix B.4 define verification via LLM judgments; no accuracy, agreement, calibration, or error analysis is reported.
  • domain assumption The paired bundle-level test r+ > r- is a valid operationalization of 'task genuinely depends on designated skills.'
    Tasks that fail this test are never used, so this test defines part of the training and evaluation phenomenon itself.
  • domain assumption External benchmarks measure skill-use rather than general ability, benchmark contamination, or incidental instruction-following.
    The withheld-skills comparison is the only analysis separating skill-use from generic ability; no contamination or general-capability checks are reported.
  • domain assumption Supervised fine-tuning on synthetic trajectories does not degrade general capabilities.
    No general-purpose benchmarks (e.g., MMLU, coding, math) are reported after SFT, though the agent benchmarks partially cover general task execution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation." pith.science (2026). https://pith.science/paper/SUIPLK42

@misc{pith2026260802287,
  author       = {Pith},
  title        = {Pith review of: SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SUIPLK42}},
  note         = {Machine review of arXiv:2608.02287}
}
read the original abstract

Agent skills have become an important mechanism for equipping language-model agents with reusable procedural knowledge. However, providing skills alone does not guarantee that current models can effectively identify, apply, and coordinate them. To improve skill-use capabilities, we introduce SKT, a verified data synthesis pipeline that constructs skill-grounded tasks and executable trajectories from large collections of agent skills. SKT selects suitable single-skill and multi-skill configurations, synthesizes tasks through rule-based and agent-based verification with feedback-guided repair, and retains only successful trajectories that substantially use every required skill. Using 2,000 public skills, SKT produces 4,000 task packages and 27,164 verified trajectories. Based on the same pipeline and a disjoint test pool, we further construct SkillEval, a held-out executable benchmark for evaluating skill use. Experiments across diverse models, benchmarks, and agent harnesses show that supervised fine-tuning on SKT-generated trajectories consistently improves skill-use performance. Verification ablations, cross-harness evaluation, and scaling experiments further demonstrate that these gains depend on high-quality supervision, extend beyond a single agent interface, and increase with broader skill coverage. Together, these results establish verified data synthesis as an effective and scalable approach for skill-use training.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 6 linked inside Pith

  1. [1]

    2026 , eprint =

    Zhang, Lisheng and Wang, Lilong and Sun, Xiangyu and Tang, Wei and Su, Haoyang and Qian, Yuehui and Yang, Qikui and Li, Qingsong and Tang, Zhenyu and Sun, Haoran and Han, Yingnan and Jiang, Yankai and Lou, Wenjie and Zhou, Bowen and Wang, Xiaosong and Bai, Lei and Xie, Zhengwei , journal =. 2026 , eprint =

    2026

  2. [2]

    arXiv preprint arXiv:2603.02176 , year =

    Organizing, Orchestrating, and Benchmarking Agent Skills at Ecosystem Scale , author =. arXiv preprint arXiv:2603.02176 , year =

    arXiv

  3. [3]

    2026 , eprint =

    Li, Xiangyi and Liu, Yimin and Chen, Wenbo and others , journal =. 2026 , eprint =

    2026

  4. [4]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  5. [5]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  6. [6]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  7. [7]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  8. [8]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

Show all 46 references
  1. [9]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  2. [10]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  3. [11]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  4. [12]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  5. [13]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  6. [14]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  7. [15]

    Advances in Neural Information Processing Systems , volume =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems , volume =

  8. [16]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =

  9. [17]

    Advances in Neural Information Processing Systems , volume =

    Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems , volume =

  10. [18]

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and Zhang, Shudan and Deng, Xiang and Zeng, Aohan and Du, Zhengxiao and Zhang, Chenhui and Shen, Sheng and Zhang, Tianjun and S...

  11. [19]

    2025 , month = oct, howpublished =

    Equipping Agents for the Real World with Agent Skills , author =. 2025 , month = oct, howpublished =

    2025

  12. [20]

    arXiv preprint arXiv:2512.17102 , year =

    Reinforcement Learning for Self-Improving Agent with Skill Library , author =. arXiv preprint arXiv:2512.17102 , year =. 2512.17102 , archivePrefix =

    arXiv

  13. [21]

    2026 , eprint =

    Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and Zheng, Zeyu and Xie, Cihang and Yao, Huaxiu , journal =. 2026 , eprint =

    2026

  14. [22]

    2026 , eprint =

    Shi, Yaorui and Chen, Yuxin and Lu, Zhengxi and Miao, Yuchun and Liu, Shugui and Gu, Qi and Cai, Xunliang and Wang, Xiang and Zhang, An , journal =. 2026 , eprint =

    2026

  15. [23]

    2026 , eprint =

    Lu, Zhengxi and Yao, Zhiyuan and Wu, Jinyang and Han, Chengcheng and Gu, Qi and Cai, Xunliang and Lu, Weiming and Xiao, Jun and Zhuang, Yueting and Shen, Yongliang , journal =. 2026 , eprint =

    2026

  16. [24]

    2026 , eprint =

    Lin, Hongxiang and Kuai, Zhirui and Xue, Erpeng and Wang, Lei , journal =. 2026 , eprint =

    2026

  17. [25]

    2026 , eprint =

    Zhu, Jiapeng and Yu, Jianxiang and Zhao, Yibo and Han, Chengcheng and Gu, Qi and Cai, Xunliang and Li, Xiang and Qian, Weining , journal =. 2026 , eprint =

    2026

  18. [26]

    2026 , eprint =

    He, Zhongyu and Li, Yuanfan and Huang, Fei and Chen, Tianyu and Chen, Siyuan and Li, Xingyang and Yu, Meng Hsuan and Liu, Xiangrong and Wei, Leyi and Pan, Lu and Zeng, Ke and Cai, Xunliang , journal =. 2026 , eprint =

    2026

  19. [27]

    and Shang, Jingbo and McAuley, Julian and Wu, Junda , journal =

    Vishe, Yash and Surana, Rohan and Jiang, Xunyi and Huang, Zihan and Li, Xintong and Kuang, Nikki Lijing and Yu, Tong and Rossi, Ryan A. and Shang, Jingbo and McAuley, Julian and Wu, Junda , journal =. 2026 , eprint =

    2026

  20. [28]

    Skill-to-

    Zhang, Tianyi and Qi, Zhonghao , journal =. Skill-to-. 2026 , eprint =

    2026

  21. [29]

    2026 , eprint =

    Yu, Aofan and Zhou, Chenyu and Xu, Tianyi and Guo, Zihan and Shan, Rong and Fu, Zhihui and Wang, Jun and Liu, Weiwen and Yu, Yong and Zhang, Weinan and Lin, Jianghao , journal =. 2026 , eprint =

    2026

  22. [30]

    arXiv preprint arXiv:2606.30015 , year =

    Parametric Skills , author =. arXiv preprint arXiv:2606.30015 , year =. 2606.30015 , archivePrefix =

    arXiv

  23. [31]

    2026 , eprint =

    Han, Tingxu and Zhang, Yi and Song, Wei and Fang, Chunrong and Chen, Zhenyu and Sun, Youcheng and Hu, Lijie , journal =. 2026 , eprint =

    2026

  24. [32]

    Towards Compositional Generalization of

    Wei, Yifan and Du, Li and Yu, Xiaoyan and Feng, Yang and Li, Angsheng , journal =. Towards Compositional Generalization of. 2026 , eprint =

    2026

  25. [33]

    and Wang, Yunhong , journal =

    Cheng, Zihao and Wang, Hongru and Liu, Zeming and Wang, Xinyi and Zhu, Xiangrong and Guo, Yuhang and Lin, Wei and Pan, Jeff Z. and Wang, Yunhong , journal =. 2026 , eprint =

    2026

  26. [34]

    2026 , url =

    Xie, Jingxu and Xu, Dylan and Zhao, Xuandong and Song, Dawn , booktitle =. 2026 , url =

    2026

  27. [35]

    arXiv preprint arXiv:2604.25727 , year =

    Toward Scalable Terminal Task Synthesis via Skill Graphs , author =. arXiv preprint arXiv:2604.25727 , year =. 2604.25727 , archivePrefix =

    arXiv

  28. [36]

    2026 , eprint =

    Zhu, Jiayin and Mao, Kelong and Guo, Yudong and He, Dengbo and Xu, Sulong and Gu, Simiu and Yue, Yutao , journal =. 2026 , eprint =

    2026

  29. [37]

    2026 , eprint =

    Zhu, Kaijie and Nie, Yuzhou and Li, Yijiang and Huang, Yiming and Wu, Jialian and Liu, Jiang and Sun, Ximeng and Yin, Zhenfei and Wang, Lun and Liu, Zicheng and Barsoum, Emad and Wang, William Yang and Guo, Wenbo , journal =. 2026 , eprint =

    2026

  30. [38]

    2026 , eprint =

    Hua, Zhanbo and Yao, Yifan and Xie, Weihao and Zhao, Yongchi and Liu, Minghao and Qiu, Ruizhi and Huang, Zhewei and Wang, Zun and Ji, Yiyan and Ye, Yunhai and Zhu, Letian and Lei, Xinping and Li, Han and Ma, Zhiyuan and Wang, Zili and Zhang, Zhaoxiang and Liu, Jiaheng , journa...

    2026

  31. [39]

    2024 , address =

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan , booktitle =. 2024 , address =. doi:10.18653/v1/2024.acl-demos.38 , url =

    2024 doi

  32. [40]

    2607.02770 , archivePrefix =

    arXiv preprint arXiv:2607.02770 , year =. 2607.02770 , archivePrefix =

    arXiv

  33. [41]

    Introducing

    Qu, Andrew , year =. Introducing

  34. [42]

    2026 , howpublished =

    2026

  35. [43]

    2606.19348 , archivePrefix =

    arXiv preprint arXiv:2606.19348 , year =. 2606.19348 , archivePrefix =

  36. [44]

    2025 , howpublished =

    2025

  37. [45]

    arXiv preprint arXiv:2605.26494 , year =

    The. arXiv preprint arXiv:2605.26494 , year =. 2605.26494 , archivePrefix =

    arXiv

  38. [46]

    2602.15763 , archivePrefix =

    arXiv preprint arXiv:2602.15763 , year =. 2602.15763 , archivePrefix =

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.