Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Skill-SP claims that an evolving library of modular skills lets LLM self-play generate diverse, verifiable tasks, outperforming unguided self-play on tool use and reasoning.

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 · deepseek-v4-flash

2026-08-01 04:24 UTC pith:XUHD7XK7

load-bearing objection A real co-evolutionary mechanism with consistent gains, but the tool-call validity gate is self-referential and needs failure analysis before the headline numbers are trusted. the 4 major comments →

arxiv 2607.22529 v1 pith:XUHD7XK7 submitted 2026-07-24 cs.CL

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

Siyuan Huang , Pengyu Cheng , Haotian Liu , Tao Chen , Yihao Liu , Jingwei Ni , Shijie Zhou , Ziyi Yang
show 5 more authors
Gangwei Jiang Mengyu Zhou Yu Cheng Xiaoxi Jiang Guanjun Jiang
This is my paper
classification cs.CL
keywords self-playskill libraryLLM self-evolutioncurriculum learningtool callinglogical reasoningreinforcement learningsynthetic data verification
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.

The paper addresses a known tension in LLM self-evolution: tasks verified by fixed environments are reliable but narrow, while open-ended task generation is broad but noisy, and passive filtering lets errors accumulate. Its proposal, Skill-SP, makes the skill—a reusable package of generation rules, examples, and executable validators—the unit that steers the whole loop. A proposer writes tasks conditioned on sampled skills, a solver trains on tasks at its current difficulty frontier, and a controller refines, prunes, and induces skills from execution feedback. On tool calling and logical reasoning across five open-weight backbones, the loop is reported to beat unguided self-play, with gains up to +42.9 points on tool use and +12.0 on reasoning, including turnarounds for models that initially could not produce valid tasks at all. The upshot: verifiable, open-ended self-improvement may be achievable without an external environment or a stronger teacher.

Core claim

Skill-SP's central claim is that a skill library can act as a proactive task-pattern interface, not a passive post-hoc filter. Each skill bundles routing metadata, procedural rules, generation hints, few-shot examples, executable validators, and usage statistics; the proposer samples skills to generate candidate tasks, complements them with an unguided exploration stream, and only tasks passing schema, contract, and probe-consistency checks enter the solver's frontier-ranked curriculum. The controller then evolves the library—refining packages from failure traces, pruning saturated ones, and inducing new ones from successful open-ended exploration. The reported consequence is that skill-rout

What carries the argument

The load-bearing object is the skill package s=⟨m,r,h,e,ν,σ⟩, a modular, machine-readable bundle of routing metadata, procedural rules, hints, few-shot examples, executable validators, and historical statistics. It does three jobs at once: injecting structural priors before the proposer writes a task, providing executable checks after the solver responds, and tracking difficulty so the controller can reroute, refine, prune, or induce skills. Around it, the loop is driven by a gated frontier reward—validity mask times a medium-difficulty score—and a bi-level update: the proposer maximizes that reward against the current solver, the solver maximizes environment reward on the top-ranked accepte

Load-bearing premise

The load-bearing premise is that a task is 'valid' when K rollouts of the current solver give a unique majority answer matching the proposer's reference; if the model shares systematic biases, all K rollouts can confidently reproduce the same wrong answer and the filter will certify it as valid.

What would settle it

Take K=10 probe rollouts from a random sample of accepted Skill-SP tasks and compare the majority answer to an independent oracle (human-verified reference or deterministic solver). If more than ~5% of certified references are wrong, the filter is validating model self-consistency, not task correctness.

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

If this is right

  • If the reported gains hold, self-improvement no longer needs a fixed environment: any domain with a machine-readable verification contract (tool schemas, constraint solvers) can be organized as an evolving skill library.
  • Competent backbones gain steadily (roughly +2.8 to +6.5 points on tool use, +1.0 to +12.0 on reasoning), while initially misaligned models can jump tens of points on tool use, because skills give them a way to synthesize valid tasks at all.
  • The ablation pattern implies that both the co-evolution of proposer and solver and the dynamic updating of skills are necessary; freezing any one component costs 2–3 points, and static skills barely beat unguided self-play.
  • Curriculum quality improves along two measurable axes at once: tasks stay near the frontier (mean solver success near 0.57) and the pool spreads out in embedding space, whereas unguided self-play collapses to narrow regions and cannot bootstrap logical puzzle generation.
  • The same recipe should extend to code generation, planning, and other contract-verifiable tasks, because the skill package only assumes a task can be represented as a prompt plus a hidden machine-checkable contract.

Where Pith is reading between the lines

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

  • A testable extension: audit the probe-consistency gate against an external oracle. If a substantial share of accepted tasks have the wrong reference answer, the gate is certifying model bias, and the reported benchmark gains would shrink out-of-distribution.
  • Because the initial library is shared across all backbones and most of the gain comes from online evolution, transferring an evolved library from a strong model to a weak one could bootstrap models that cannot currently generate valid puzzles; the paper's limitations section points in this direction.
  • The dual-stream design suggests a general engineering recipe: keep an exploration stream always running and absorb its successes into the skill library; this is a cheap way to keep any self-play loop from mode-collapsing, independent of the specific reward.
  • One could quantify the marginal value of each newly induced skill by deleting it from the library and measuring the change in accepted task diversity and accuracy; the paper reports package counts but not per-skill attribution.

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

4 major / 5 minor

Summary. The paper proposes Skill Self-Play (Skill-SP), a self-evolution framework with three components: a proposer that generates tasks, a solver that learns to solve them, and a controller that maintains an evolving library of skill packages. The skill library guides task generation via routing and provides structural validators, while the proposer is trained to target the solver's learning frontier and the solver is trained on a mixed curriculum of skill-guided and open-ended tasks. The framework is evaluated on tool-call prediction (API-Bank, four BFCL categories) and logical reasoning (ZebraLogic) across five open-source backbones. The reported results show consistent gains over the base models and over an 'Unguided SP' baseline, with especially large gains on initially weak Ministral models. Ablations on Qwen3-4B-Instruct decompose the contributions of skill routing, skill evolution, and co-evolutionary updates. Diagnostics report that skill-routed tasks are near the solver's frontier, that the task pool is diverse, and that the skill library grows over iterations.

Significance. If the reported effects are robust, Skill-SP is a meaningful advance: it offers a concrete mechanism for combining structured verification with open-ended task diversity, and it demonstrates large turnaround effects on initially misaligned models. The paper includes detailed package-level case studies, a deterministic logical-reasoning checker, and a public code release, which are strengths for reproducibility. The work is also relevant to the growing literature on self-evolving LLMs. However, the empirical case is currently incomplete: all reported numbers come from single training runs, ablations use one backbone, the BFCL evaluation covers a selected subset of categories, and the main comparison is only to an in-house 'Unguided SP' baseline rather than to existing self-evolution methods. The circularity of the frontier-proximity diagnostic and the lack of failure analysis for the self-consistency validity gate are additional concerns. These issues do not invalidate the central idea, but they need to be addressed before the strength of the paper's claims can be fully assessed.

major comments (4)
  1. [§3.2, Eq. (4); Appendix C] The 'probe consistent' validity gate is load-bearing for the tool-calling results, yet it is never checked against an external oracle. Accepting a task when K rollouts of the current solver agree with the proposer-defined reference can certify systematically wrong tool calls whenever the proposer and solver share the same model-wide biases. The paper reports no false-acceptance rate, no comparison against a stronger or independent verifier, and no analysis of the failure modes of this gate. Since the headline +42.9-point Ministral-3-8B gain (Table 1) is driven by this self-referential filter, the manuscript should provide at least a sampled audit (e.g., external judge or deterministic schema/argument checks) and report disagreement rates between the probe-consistency verdict and an independent oracle.
  2. [§4.4, Fig. 5a; Appendix C] The 'frontier proximity' diagnostic (mean v_solve ≈ 0.57 for the skill stream) is partly an artifact of the method. Appendix C states that candidates are retained only when v_solve ∈ [0.25, 0.75], and Eq. (2) rewards v_solve near 0.5. A mean of 0.57 is therefore close to what the filter itself enforces, not independent evidence that the skill library anchors tasks at the frontier. To make the diagnostic meaningful, report the pre-filter v_solve distribution, acceptance rates, and a comparison against tasks sampled uniformly from the filtering interval. This does not undo the benchmark gains, but it removes a currently circular piece of evidence.
  3. [§4.2, Tables 1–2; §4.3, Tables 3–4] All reported results are single training runs without confidence intervals, significance tests, or multiple seeds. Many of the claimed gains are small (e.g., +0.8 on BFCL for Qwen3-8B, +0.5 on Java for Granite-4.1-3B in Table 1), and the ablations are performed on only one backbone (Qwen3-4B-Instruct). Given the paper's central claim that Skill-SP 'consistently' pushes the performance ceiling, the absence of variance information is a substantive gap. At minimum, repeat the main comparisons and ablations with 2–3 seeds and report means and standard errors, or justify why single-run results are conclusive for the stated effect sizes.
  4. [§4.1; §4.2; Appendix G] The evaluation scope weakens the external-validity claim. BFCL is represented by only four of its categories (simple_javascript, simple_python, simple_java, and live_simple), and no full-BFCL or other held-out tool-calling benchmark is reported. More importantly, the only self-evolution baseline is the in-house 'Unguided SP', and it is absent from the reasoning domain because it fails to bootstrap. This makes it hard to judge how Skill-SP compares with existing self-evolution methods (e.g., self-play fine-tuning, SPIN-style methods, or other skill-based self-evolution pipelines). The paper should either add a stronger published baseline or explicitly position the contribution as 'skill-guidance over unguided self-play' and temper the general 'robust evolution engine' claim accordingly.
minor comments (5)
  1. [Abstract and §5] Phrases like 'fundamental paradigm shift' and 'truly open-ended frontier' overstate the evidence. The paper itself (Appendix H) acknowledges fixed heuristics and minimal-capability requirements, so the conclusion should be more guarded.
  2. [§4.4, Fig. 5b] The diversity diagnostic uses PCA visualization but no quantitative measure of coverage. Reporting a quantitative diversity score (e.g., average pairwise distance or coverage in embedding space) would make the claim easier to verify.
  3. [Table 3, 'Frozen skills'] The 'Frozen skills' variant is described as holding the initial library fixed, but the terminology could be confused with frozen solver/proposer variants in Table 4. Consider renaming to 'Static skill library' for clarity.
  4. [Appendix H] The limitation statement that the curriculum blending ratio α and difficulty bounds 'may require empirical tuning for new task families' is important. It should be surfaced in the main paper, not only in a supplementary appendix, because it directly qualifies the 'autonomous' claim.
  5. [§4.1, BFCL subset] The choice of four BFCL categories should be justified in the main text. BFCL has many categories; selective reporting could inflate or deflate average gains depending on which categories are hardest for the base models.

Circularity Check

1 steps flagged

Frontier-proximity diagnostic is partly by construction, but benchmark gains are externally grounded.

specific steps
  1. fitted input called prediction [Section 4.4 (Data-Loop Diagnostics, Figure 5a); Section 3.1 Eq. 2; Appendix C Table 5]
    "Curriculum Quality. Figure 5a tracks the empirical solver success rate 𝑣solve. The skill-routed stream consistently generates tasks closer to the solver's learning frontier with a mean 𝑣solve of approximately 0.57. ... Candidate filtering: Probe each candidate with 𝐾=10 frozen-solver responses; retain records with 𝑣solve(𝒙,𝒄;𝜋(𝑡)solve) ∈ [0.25, 0.75]"

    The 'learning frontier' is operationalized by the proposer reward Rpropose = 𝟙{valid}·(1−2|vsolve−0.5|) (Eq. 2), and the candidate filter retains only records with vsolve ∈ [0.25,0.75]. The reported mean vsolve ≈ 0.57 is therefore a direct arithmetic consequence of the acceptance window and the reward's peak at 0.5, not an independent empirical discovery. The paper presents this as evidence that Skill-SP 'anchors task generation at the model's evolving learning frontier,' but the diagnostic reduces to the selection rule that defines the training pool. This is a self-referential internal validation, though it does not bear on the external benchmark results.

full rationale

The central empirical claim—Skill-SP improves held-out tool-use and logical-reasoning performance relative to base models and Unguided SP—is measured on external benchmarks (API-Bank, BFCL, ZebraLogic) with fixed human-constructed reference answers. Those scores do not reduce to the self-play loop's internal rewards, validity filters, or skill-library statistics. The only identifiable circularity is the internal 'frontier proximity' diagnostic (Figure 5a): because the training pool is explicitly filtered to vsolve ∈ [0.25,0.75] and ranked by 1−2|vsolve−0.5|, the observed mean vsolve ≈ 0.57 is substantially by construction. This is a partial, non-load-bearing circularity; the headline benchmark gains remain independent of it. Self-citations in the related work are contextual and not used to justify the framework's core premise, and no uniqueness theorem or ansatz is imported via self-citation. The probe-consistency gate (Eq. 4) poses a potential correctness/fidelity risk—shared model biases could certify consistent-but-wrong references—but this is a validity concern about the training signal, not a circular derivation of the reported held-out results. Accordingly, the overall circularity score is low.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 1 invented entities

The central claim is an empirical training-loop claim. Its main hidden inputs are hand-set curriculum hyperparameters (α, [0.25,0.75], K, routing weights, thresholds γ), the unverified validity-by-self-consistency assumption, and the assumption that self-generated verified data transfers to benchmarks. No new physical entities are posited; the skill package is a software abstraction.

free parameters (7)
  • Curriculum blending ratio α = 0.5
    Eq. 6 splits the training pool equally between skill-stream and exploration-stream records; no sensitivity analysis is reported.
  • Frontier difficulty bounds (v_solve range) = [0.25, 0.75]
    Candidate filtering retains tasks whose estimated solver success is in [0.25,0.75]; this defines 'frontier' and shapes every curriculum (Appendix C).
  • Probe count K = 10
    v_solve is estimated from K=10 frozen-solver rollouts (Eq. 1, Appendix C); the noise of this estimate is not quantified.
  • Skill routing weights and exploration bonus = κ=4; w_min=0.25, w_max=1.75; β=1; τ=8
    Eqs. 11-13 set skill sampling probabilities; these hand-chosen values control which skills are exploited/explored.
  • Reward decomposition weights = 0.9 exact-match / 0.1 format
    Eqs. 14-15 weight exact correctness against format compliance for both domains.
  • Skill pruning/induction thresholds γprune, γinduce = not reported
    Eqs. 8-9 trigger skill retirement and induction; the paper never reports their numerical values, which is a reproducibility gap.
  • GRPO optimization schedule = 5/15 proposer/solver steps (tool), 3/10 (reasoning); lr 1e-6
    Training dynamics and final accuracy depend on these hand-set update counts and learning rate; no ablation on compute.
axioms (4)
  • domain assumption Task validity equals schema compliance ∧ contract validity ∧ probe consistency, where probe consistency is majority agreement of K solver rollouts with the proposer's reference.
    Eq. 4 defines validity; this assumes self-consistency implies correctness although proposer and solver share initialization and can share biases.
  • domain assumption Training on self-generated, machine-verified tasks transfers to held-out benchmarks.
    No theoretical transfer guarantee; the entire empirical claim rests on this.
  • domain assumption The skill tuple s=⟨m,r,h,e,ν,σ⟩ is a sufficient interface for reusable task patterns, and the initial base policy can induce new integrity- and novelty-filtered skills from exploration traces.
    Eq. 10 and Appendix D assume π_control can abstract useful packages; induction quality is only shown via two case studies.
  • standard math GRPO optimization of proposer/solver remains stable and rewarding at the five-iteration horizon.
    GRPO is a standard algorithm (Shao et al., 2024) assumed to optimize the defined objectives without catastrophic reward hacking beyond the gating filter.
invented entities (1)
  • Evolving skill library (skill package interface) no independent evidence
    purpose: Serves as a training-time abstraction that routes task generation, provides executable validators, and stores reusable task patterns.
    This is a software abstraction introduced by the paper rather than a physical entity; ablations provide internal support, but there is no falsifiable handle outside this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 28984 in / 15979 out tokens · 157198 ms · 2026-08-01T04:24:22.814977+00:00 · methodology

0 comments
read the original abstract

LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound methods obtain precise feedback but confine learning to narrow domains, while open-ended self-generation broadens the task space but lacks reliable verification, allowing misleading rewards to pollute the training loop. We identify agent skills as a powerful middle ground to reconcile this tension: each skill ensures deep, verifiable execution in a specific scenario, while dynamic routing across skills maintains open-ended task variety. Leveraging this insight, we introduce Skill Self-Play (Skill-SP), a co-evolutionary framework comprising a proposer, a solver, and a dynamic skill controller. Orchestrated via a reinforcement learning loop, these components co-evolve in a continuous self-play loop: the proposer generates challenging tasks conditioned on dynamically sampled skills; the solver explores candidate solutions to push its capability boundaries; and the skill controller collects execution feedback to update and expand the skill library. This interactive co-evolution effectively bridges the gap between structured verification and open-ended exploration. Empirical evaluations on tool-use and reasoning benchmarks demonstrate that Skill-SP, serving as a robust evolution engine, consistently pushes the performance ceiling of competent backbones while catalyzing striking turnarounds for initially misaligned models. Our code is available at https://github.com/Qwen-Applications/skill-self-play.

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

45 extracted references · 27 linked inside Pith

  1. [1]

    Ultraif: Advancing instruction following from the wild

    Kaikai An, Li Sheng, Ganqu Cui, Shuzheng Si, Ning Ding, Yu Cheng, and Baobao Chang. Ultraif: Advancing instruction following from the wild. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 18722–18737,

    2025

  2. [3]

    Skilljuror: Measuring how agent skill organization changes runtime behavior.arXiv preprint arXiv:2606.11543, 2026b

    Zhiyu Chen, Zihan Guo, Bo Huang, Bingwei Lu, Jianghao Lin, Yuanjian Zhou, and Weinan Zhang. Skilljuror: Measuring how agent skill organization changes runtime behavior.arXiv preprint arXiv:2606.11543, 2026b. 10 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu...

    Pith/arXiv arXiv

  3. [4]

    I want to plan a dinner for my friends this weekend in Portland. We’re looking for a nice Italian place that doesn’t cost more than $30 per person

    23 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Proposer Construction Rules(𝑟)skill_038 Trigger The user requests a table booking after a restaurant discovery and specifies a date, time, and guest count. Required evidence A prior observation contains a top-rated restaurant with an identifier, cuisine, and price constrain...

    2023

  4. [6]

    From self-evolving synthetic data to verifiable-reward rl: Post-training multi-turn interactive tool-using agents

    Jiaxuan Gao, Jiaao Chen, Chuyi He, Wei-Chen Wang, Shusheng Xu, Hanrui Wang, Di Jin, and Yi Wu. From self-evolving synthetic data to verifiable-reward rl: Post-training multi-turn interactive tool-using agents. arXiv preprint arXiv:2601.22607, 2026a. Yudong Gao, Zongjie Li, Zimo Ji, Pingchuan Ma, Shuai Wang, et al. Skillreducer: Optimizing llm agent skills...

    arXiv

  5. [7]

    Rlef: Grounding code llms in execution feedback with reinforcement learning.arXiv preprint arXiv:2410.02089,

    Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Quentin Carbonneaux, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code llms in execution feedback with reinforcement learning.arXiv preprint arXiv:2410.02089,

    Pith/arXiv arXiv

  6. [8]

    Gems: Agent-native multimodal generation with memory and skills.arXiv preprint arXiv:2603.28088,

    Zefeng He, Siyuan Huang, Xiaoye Qu, Yafu Li, Tong Zhu, Yu Cheng, and Yang Yang. Gems: Agent-native multimodal generation with memory and skills.arXiv preprint arXiv:2603.28088,

    arXiv

  7. [9]

    R-zero: Self-evolving reasoning llm from zero data.arXiv preprint arXiv:2508.05004, 2025a

    Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, Ruosen Li, Jiaxin Huang, Haitao Mi, and Dong Yu. R-zero: Self-evolving reasoning llm from zero data.arXiv preprint arXiv:2508.05004, 2025a. Siyuan Huang, Xiaoye Qu, Yafu Li, Yun Luo, Zefeng He, Daizong Liu, and Yu Cheng. Spotlight on token perception for multimodal reinforcement learni...

    Pith/arXiv arXiv

  8. [10]

    Bowen Jiang, Taiwei Shi, Ryo Kamoi, Yuan Yuan, Camillo J Taylor, Longqi Yang, Pei Zhou, and Sihao Chen

    Accessed: 2026-04-28. Bowen Jiang, Taiwei Shi, Ryo Kamoi, Yuan Yuan, Camillo J Taylor, Longqi Yang, Pei Zhou, and Sihao Chen. One model, all roles: Multi-turn, multi-agent self-play reinforcement learning for conversational social intelligence. arXiv preprint arXiv:2602.03109, 2026a. Dongwei Jiang, Jingyu Zhang, Orion Weller, Nathaniel Weir, Benjamin Van ...

    arXiv 2026

  9. [11]

    Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al

    URLhttps://openreview.net/forum? id=buwLCdOHxO. Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines.arXiv preprint arXiv:2310.03714,

    Pith/arXiv arXiv

  10. [13]

    R-diverse: Mitigating diversity illusion in self-play llm training.arXiv preprint arXiv:2602.13103, 2026a

    Gengsheng Li, Jinghan He, Shijie Wang, Dan Zhang, Ruiqi Liu, Renrui Zhang, Zijun Yao, Junfeng Fang, Haiyun Guo, and Jinqiao Wang. R-diverse: Mitigating diversity illusion in self-play llm training.arXiv preprint arXiv:2602.13103, 2026a. Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank:...

    arXiv 2023

  11. [14]

    Skillsbench: Benchmarking how well agent skills work across diverse tasks.arXiv preprint arXiv:2602.12670, 2026b

    Xiangyi Li, Wenbo Chen, Yimin Liu, Shenghan Zheng, Xiaokun Chen, Yifeng He, Yubo Li, Bingran You, Haotian Shen, Jiankai Sun, et al. Skillsbench: Benchmarking how well agent skills work across diverse tasks.arXiv preprint arXiv:2602.12670, 2026b. Yanchao Li, Wanhao Liu, Ben Gao, Jiaqing Xie, Zhehong Ai, Na Zou, Yuqiang Li, and Tianfan Fu. Skillsinjector: D...

    Pith/arXiv arXiv

  12. [15]

    Agent skills: A data-driven analysis of claude skills for extending large language model functionality.arXiv preprint arXiv:2602.08004,

    George Ling, Shanshan Zhong, and Richard Huang. Agent skills: A data-driven analysis of claude skills for extending large language model functionality.arXiv preprint arXiv:2602.08004,

    arXiv

  13. [16]

    Ministral 3.arXiv preprint arXiv:2601.08584, 2026a

    Alexander H Liu, Kartik Khandelwal, Sandeep Subramanian, Victor Jouault, Abhinav Rastogi, Adrien Sadé, Alan Jeffares, Albert Jiang, Alexandre Cahill, Alexandre Gavaudan, et al. Ministral 3.arXiv preprint arXiv:2601.08584, 2026a. Bo Liu, Leon Guertler, Simon Yu, Zichen Liu, Penghui Qi, Daniel Balcells, Mickel Liu, Cheston Tan, Weiyan Shi, Min Lin, et al. S...

    Pith/arXiv arXiv 2024

  14. [17]

    Skillrevise: Improving llm-authored agent skills via trace-conditioned skill revision

    Yuxuan Liu, Zhaochen Su, Lingyun Xie, Yuhao Zhang, Qing Zong, Jiahe Guo, Zhongwei Xie, Yiyan Ji, Yauwai Yim, Hongyu Luo, et al. Skillrevise: Improving llm-authored agent skills via trace-conditioned skill revision. arXiv preprint arXiv:2606.01139, 2026d. Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. On llms-driven syn...

    Pith/arXiv arXiv 2024

  15. [18]

    Search self-play: Pushing the frontier of agent capability without supervision.arXiv preprint arXiv:2510.18821,

    Hongliang Lu, Yuhang Wen, Pengyu Cheng, Ruijin Ding, Jiaqi Guo, Haotian Xu, Chutian Wang, Haonan Chen, Xiaoxi Jiang, and Guanjun Jiang. Search self-play: Pushing the frontier of agent capability without supervision.arXiv preprint arXiv:2510.18821,

    Pith/arXiv arXiv

  16. [19]

    Skill0: In-context agentic reinforcement learning for skill internalization.arXiv preprint arXiv:2604.02268,

    Zhengxi Lu, Zhiyuan Yao, Jinyang Wu, Chengcheng Han, Qi Gu, Xunliang Cai, Weiming Lu, Jun Xiao, Yueting Zhuang, and Yongliang Shen. Skill0: In-context agentic reinforcement learning for skill internalization.arXiv preprint arXiv:2604.02268,

    Pith/arXiv arXiv

  17. [20]

    Procmem: Learning reusable procedural memory from experience via non-parametric ppo for llm agents.arXiv preprint arXiv:2602.01869,

    Qirui Mi, Zhijian Ma, Mengyue Yang, Haoxuan Li, Yisen Wang, Haifeng Zhang, and Jun Wang. Procmem: Learning reusable procedural memory from experience via non-parametric ppo for llm agents.arXiv preprint arXiv:2602.01869,

    Pith/arXiv arXiv

  18. [21]

    Better alignment with instruction back-and-forth translation

    Thao Nguyen, Jeffrey Li, Sewoong Oh, Ludwig Schmidt, Jason E Weston, Luke Zettlemoyer, and Xian Li. Better alignment with instruction back-and-forth translation. InFindings of the Association for Computational Linguistics: EMNLP 2024, pp. 13289–13308,

    2024

  19. [22]

    Trace2skill: Distill trajectory-local lessons into transferable agent skills.arXiv preprint arXiv:2603.25158,

    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,

    Pith/arXiv arXiv

  20. [23]

    Largelanguagemodelscanself-improveatwebagenttasks.arXivpreprintarXiv:2405.20309,

    Ajay Patel, Markus Hofmarcher, Claudiu Leoveanu-Condrei, Marius-Constantin Dinu, Chris Callison-Burch, and SeppHochreiter. Largelanguagemodelscanself-improveatwebagenttasks.arXivpreprintarXiv:2405.20309,

    Pith/arXiv arXiv

  21. [24]

    Infobench: Evaluating instruction following ability in large language models

    Yiwei Qin, Kaiqiang Song, Yebowen Hu, Wenlin Yao, Sangwoo Cho, Xiaoyang Wang, Xuansheng Wu, Fei Liu, Pengfei Liu, and Dong Yu. Infobench: Evaluating instruction following ability in large language models. In Findings of the Association for Computational Linguistics: ACL 2024, pp. 13025–13048,

    2024

  22. [25]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11

    2019

  23. [27]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  24. [28]

    Skill- foundry: Building self-evolving agent skill libraries from heterogeneous scientific resources.arXiv preprint arXiv:2604.03964,

    Shuaike Shen, Wenduo Cheng, Mingqian Ma, Alistair Turcan, Martin Jinye Zhang, and Jian Ma. Skill- foundry: Building self-evolving agent skill libraries from heterogeneous scientific resources.arXiv preprint arXiv:2604.03964,

    Pith/arXiv arXiv

  25. [30]

    A survey on self-evolution of large language models.arXiv preprint arXiv:2404.14387,

    Zhengwei Tao, Ting-En Lin, Xiancai Chen, Hangyu Li, Yuchuan Wu, Yongbin Li, Zhi Jin, Fei Huang, Dacheng Tao, and Jingren Zhou. A survey on self-evolution of large language models.arXiv preprint arXiv:2404.14387,

    Pith/arXiv arXiv

  26. [31]

    Code-a1: Adversarial evolving of code llm and test llm via reinforcement learning.arXiv preprint arXiv:2603.15611, 2026a

    Aozhe Wang, Yuchen Yan, Nan Zhou, Zhengxi Lu, Weiming Lu, Jun Xiao, Yueting Zhuang, and Yongliang Shen. Code-a1: Adversarial evolving of code llm and test llm via reinforcement learning.arXiv preprint arXiv:2603.15611, 2026a. Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, et al. ...

    arXiv

  27. [32]

    Toward training superintelligent software agents through self-play swe-rl

    Yuxiang Wei, Zhiqing Sun, Emily McMilin, Jonas Gehring, David Zhang, Gabriel Synnaeve, Daniel Fried, Lingming Zhang, and Sida Wang. Toward training superintelligent software agents through self-play swe-rl. arXiv preprint arXiv:2512.18552,

    Pith/arXiv arXiv

  28. [33]

    Propose, solve, verify: Self-play through formal verification.arXiv preprint arXiv:2512.18160,

    14 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Alex Wilf, Pranjal Aggarwal, Bryan Parno, Daniel Fried, Louis-Philippe Morency, Paul Pu Liang, and Sean Welleck. Propose, solve, verify: Self-play through formal verification.arXiv preprint arXiv:2512.18160,

    arXiv

  29. [34]

    Wizardlm: Empowering large pre-trained language models to follow complex instructions

    Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow complex instructions. In International Conference on Learning Representations, volume 2024, pp. 30745–30766,

    2024

  30. [35]

    Genius: A generalizable and purely unsupervised self-training framework for advanced reasoning

    Fangzhi Xu, Hang Yan, Chang Ma, Haiteng Zhao, Qiushi Sun, Kanzhi Cheng, Junxian He, Jun Liu, and Zhiyong Wu. Genius: A generalizable and purely unsupervised self-training framework for advanced reasoning. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 13153–13167, 2025a. Zhangchen Xu...

    Pith/arXiv arXiv 2025

  31. [36]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025a. Haoyan Yang, Mario Xerri, Solha Park, Huajian Zhang, Yiyang Feng, Sai Akhil Kogilathota, and Jiawei Zhou. Self-improvement of large language models: A technical overvie...

    Pith/arXiv arXiv

  32. [37]

    Cot-self-instruct: Building high-quality synthetic prompts for reasoning and non-reasoning tasks.arXiv preprint arXiv:2507.23751, 2025a

    Ping Yu, Jack Lanchantin, Tianlu Wang, Weizhe Yuan, Olga Golovneva, Ilia Kulikov, Sainbayar Sukhbaatar, Jason Weston, and Jing Xu. Cot-self-instruct: Building high-quality synthetic prompts for reasoning and non-reasoning tasks.arXiv preprint arXiv:2507.23751, 2025a. Wenhao Yu, Zhenwen Liang, Chengsong Huang, Kishan Panaganti, Tianqing Fang, Haitao Mi, an...

    Pith/arXiv arXiv

  33. [38]

    Learning to reason without external rewards.arXiv preprint arXiv:2505.19590,

    Xuandong Zhao, Zhewei Kang, Aosong Feng, Sergey Levine, and Dawn Song. Learning to reason without external rewards.arXiv preprint arXiv:2505.19590,

    Pith/arXiv arXiv

  34. [39]

    Skilllearnbench: Benchmarking continual learning methods for agent skill generation on real-world tasks.arXiv preprint arXiv:2604.20087,

    Shanshan Zhong, Yi Lu, Jingjie Ning, Yibing Wan, Lihan Feng, Yuyi Ao, Leonardo FR Ribeiro, Markus Dreyer, Sean Ammirati, and Chenyan Xiong. Skilllearnbench: Benchmarking continual learning methods for agent skill generation on real-world tasks.arXiv preprint arXiv:2604.20087,

    Pith/arXiv arXiv

  35. [40]

    Instruction-following evaluation for large language models.arXiv preprint arXiv:2311.07911,

    Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-following evaluation for large language models.arXiv preprint arXiv:2311.07911,

    Pith/arXiv arXiv

  36. [41]

    Webarena: A realistic web environment for building autonomous agents

    15 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. In International Conference on Learning Representations, volume 2024,...

    2024

  37. [42]

    Skillgenbench: Benchmarking skill generation pipelines for llm agents

    Yifan Zhou, Zhentao Zhang, Ziming Cheng, Shuo Zhang, Qizhen Lan, Zhangquan Chen, Zhi Yang, Ronghao Chen, Huacan Wang, Sen Hu, et al. Skillgenbench: Benchmarking skill generation pipelines for llm agents. arXiv preprint arXiv:2605.18693,

    Pith/arXiv arXiv

  38. [43]

    project updates

    evaluates a different verifiable agent task family: logical reasoning over grid-structured constraint puzzles. Each instance provides a natural-language puzzle whose solution is a complete assignment of attributes to houses. The evaluator checks whether the model’s output can be parsed into a valid grid, whether the predicted schema matches the reference ...

    2023

  39. [45]

    schedule a meeting

    Proposer Construction Rules(𝑟)skill_042 TriggerThe user requests a meeting between participants with a time range and topic, and an observation confirms an available time slot. Required evidence The user specifies participants, a time range (for example, 10 AM to 2 PM), and a topic; an observation confirms a particular available slot within that range. Di...

    2023

  40. [2018]

    Envscaler: Scaling tool-interactive environments for llm agent via programmatic synthesis.arXiv preprint arXiv:2601.05808,

    Xiaoshuai Song, Haofei Chang, Guanting Dong, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. Envscaler: Scaling tool-interactive environments for llm agent via programmatic synthesis.arXiv preprint arXiv:2601.05808,

    Pith/arXiv arXiv

  41. [2019]

    URLhttps://arxiv.org/abs/1908.10084. 13 Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills Qingyu Ren, Qianyu He, Jiajie Zhu, Xingzhou Chen, Jingwen Chang, Zeye Sun, Han Xia, Fei Yu, Jiaqing Liang, and Yanghua Xiao. Seif: Self-evolving reinforcement learning for instruction following.arXiv preprint arXiv:2605.07465,

    Pith/arXiv arXiv 1908

  42. [2023]

    Language self-play for data-free training.arXiv preprint arXiv:2509.07414,

    Jakub Grudzien Kuba, Mengting Gu, Qi Ma, Yuandong Tian, Vijai Mohan, and Jason Chen. Language self-play for data-free training.arXiv preprint arXiv:2509.07414,

    arXiv

  43. [2024]

    Spar: Self-play with tree-search refinement to improve instruction-following in large language models.arXiv preprint arXiv:2412.11605, 2024a

    Jiale Cheng, Xiao Liu, Cunxiang Wang, Xiaotao Gu, Yida Lu, Dan Zhang, Yuxiao Dong, Jie Tang, Hongning Wang, and Minlie Huang. Spar: Self-play with tree-search refinement to improve instruction-following in large language models.arXiv preprint arXiv:2412.11605, 2024a. Pengyu Cheng, Yong Dai, Tianhao Hu, Han Xu, Zhisong Zhang, Lei Han, Nan Du, and Xiaolong ...

    Pith/arXiv arXiv

  44. [2025]

    Skill-rm: Unifying heterogeneous evaluation criteria via agent skill.arXiv preprint arXiv:2606.03980, 2026a

    Tao Chen, Gangwei Jiang, Pengyu Cheng, Siyuan Huang, Yihao Liu, Jingwei Ni, Jiaqi Guo, Mengyu Zhou, Kai Tang, Junling Liu, et al. Skill-rm: Unifying heterogeneous evaluation criteria via agent skill.arXiv preprint arXiv:2606.03980, 2026a. Yixing Chen, Yiding Wang, Siqi Zhu, Haofei Yu, Tao Feng, Muhan Zhang, Mostofa Patwary, and Jiaxuan You. Multi-agent ev...

    Pith/arXiv arXiv

  45. [2026]

    Self- play with execution feedback: Improving instruction-following capabilities of large language models

    Guanting Dong, Keming Lu, Chengpeng Li, Tingyu Xia, Bowen Yu, Chang Zhou, and Jingren Zhou. Self- play with execution feedback: Improving instruction-following capabilities of large language models. In International Conference on Learning Representations, volume 2025, pp. 39286–39313,

    2025