Pith. sign in

REVIEW 3 major objections 5 minor 103 references

SkillBoost recasts LLM skill self-evolution as a constrained exploration–exploitation search, not a data-fitting problem, and reports that this keeps the test–train gap near zero.

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 11:40 UTC pith:4Q7TU3WK

load-bearing objection SkillBoost is a well-engineered answer to skill overfitting in self-evolving agents, but the headline near-zero overfitting claim is not independently checkable as reported. the 3 major comments →

arxiv 2607.26643 v1 pith:4Q7TU3WK submitted 2026-07-29 cs.AI cs.LG

Rethinking Self-Evolution: A Constrained Exploration-Exploitation Process for Mitigating Skill Overfitting

Hongqiang Lin , Chao Liu , Xiaofan Bai , Xuan Jin , Yuhong Li , Nenggan Zheng , Xipeng Cao This is my paper
classification cs.AI cs.LG
keywords skill self-evolutionLLM agentsskill overfittingexploration-exploitationBest-of-Nregression gateagentic skillsskill transfer
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.

This paper argues that when an LLM agent improves itself by refining reusable skills, the process should be treated as a constrained exploration–exploitation search rather than fitting the skill to observed trajectories. The proposed SkillBoost loop attributes each failure to a specific skill component, generates several repair candidates that share that diagnosis but differ in repair strategy, and commits an edit only if it improves the full split while keeping regressions below a fixed bound. The authors report consistent gains across 23 model–benchmark configurations, a near-zero test–train gap, and skills that transfer to other agents and related benchmarks. The contribution, if it holds, is a lightweight path for frozen-parameter agents to keep improving in deployment without memorizing the current batch.

Core claim

The central claim is that skill overfitting — gains on the current trajectory batch that fail to transfer — is not mainly a data-quantity problem but a consequence of optimizing the wrong objective. SkillBoost therefore casts self-evolution as a Markov decision process over skill states and runs three stages each round: structured exploitation analyzes failed trajectories to produce an evidence-grounded diagnosis naming which skill components to edit; prior-guided exploration keeps that diagnosis fixed and generates N candidate skills through different repair strategies; verified acceptance evaluates candidates on the training split and accepts only updates with r(s′)>0 and Regress(s′)<ε. Th

What carries the argument

The load-bearing mechanism is the verified-acceptance gate, Accept(s′) ⇔ r(s′)>0 ∧ Regress(s′)<ε, applied to a Best-of-N pool of candidate skills. Because r(s′) decomposes as Fix(s′)−Regress(s′), the first condition is exactly "fix more than it breaks," and the second caps how many previously correct cases may regress. This gate, together with a diagnosis step that restricts edits to failure-attributed skill components, turns skill mutation into constrained search. The theoretical complement is a Best-of-N sampling bound: with N independent candidate directions, expected per-step progress scales like ε·sqrt(2 ln N / d), so gains from enlarging the pool saturate quickly.

Load-bearing premise

The training split used both to select candidate edits and to enforce the regression bound is a sufficient stand-in for the distribution the agent will face after deployment, and the fixed ε=0.05 cap is appropriate for every task; if either fails, the near-zero test–train gap may not survive contact with new data.

What would settle it

Take a skill version accepted by the gate (r>0, Regress<0.05 on the training split) and evaluate it on a new disjoint sample from the same task family, or a lightly perturbed version of the benchmark. If the test-minus-train gap turns strongly negative, or if a held-out set's solved-case regressions exceed its fixes, then the training split is not a sufficient proxy and the claimed overfitting control does not generalize. A cheaper check: rerun the loop with ε=0 and with ε=0.05; if performance and generalization are identical, the regression bound is not doing the work claimed.

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

If this is right

  • Accepted edits are guaranteed by construction to fix more cases than they break on the training split while keeping regressions under the bound, which is what keeps the reported test–train gap near zero.
  • Skills evolve into compact procedural rules — exact function-name matching, static-value spreadsheet discipline, amnesia-robust exploration — rather than growing monotonically.
  • Best-of-N exploration shows diminishing returns as N grows; the paper finds N=4 captures most of the gain at moderate cost.
  • Optimized skills can be reused by other agents and on related benchmarks, outperforming human-written skills in the paper's transfer experiments.
  • The framework inherits the base model's ability to attribute failures, generate edits, and evaluate candidates; the paper states that weaker models may limit self-evolution quality.

Where Pith is reading between the lines

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

  • Editorial: because the acceptance gate and the selection metric use the same training split, the method's overfitting control is only as strong as that split's representativeness; adding a small held-out validation set to the acceptance rule would test this directly.
  • Editorial: the learned rules read as behavioral disciplines the model already possesses but applies unreliably zero-shot; a testable extension is whether compact versions of these rules give similar gains for smaller models or non-frontier executors.
  • Editorial: the Best-of-N theoretical model assumes independent uniform sampling, which real LLM-generated candidates violate; measuring per-round progress across N=2,4,8,16 on several tasks would show how early saturation occurs in practice.

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. The paper argues that skill self-evolution for LLM agents should be reframed as a constrained exploration–exploitation search rather than trajectory fitting. It proposes SkillBoost, a three-stage loop: structured exploitation attributes failures to specific skill components using trajectory evidence; prior-guided exploration generates N repair candidates through different repair strategies; verified acceptance commits a candidate only if it improves the full-set score and keeps regressions below a threshold. The authors report results on 23 model–benchmark configurations (20 text-based pairs plus three DocVQA pairs), component ablations, Best-of-N analyses, cross-model and cross-benchmark transfer experiments, and a theoretical lower bound on iteration complexity under a uniform-sampling model. The central claim is that SkillBoost achieves state-of-the-art performance while reducing the test–train generalization gap to near zero.

Significance. If the empirical claims hold, this is a useful contribution: it identifies skill overfitting as a distinct failure mode and shows that a generate-and-test loop with an explicit regression gate can produce large held-out gains across frontier models and diverse benchmarks. The paper's strengths include a broad evaluation across 20 model–benchmark pairs, component-level ablations, transfer tests, and a transparent acceptance criterion. The configuration-level statistical analysis is also a reasonable first step. However, the headline overfitting claim is under-supported as reported: the train accuracies behind the reported Δ values are not given, no per-run or per-question variance is reported, no code or data are released, and key hyperparameters (ε, N, T) are selected on the same benchmarks used for the main results. These gaps make the central 'near-zero overfitting' assertion not independently checkable from the manuscript.

major comments (3)
  1. [Overfitting Analysis (Table III)] The central claim that SkillBoost 'reduces overfitting to a near-zero test–train gap' rests entirely on Δ = Test − Train in Table III, but the Train values are not reported. Since D is the same split used for rollout, candidate selection, and the regression gate, the gate directly constrains the Train side of this gap, so a near-zero Δ could arise from a low Train accuracy rather than a genuinely high Test accuracy. Several SkillBoost entries are negative (e.g., SpreadsheetBench/Kimi-k2.6: −3.8; ALFWorld/Kimi-k2.6: −1.2), and the appendix's LiveMath example uses only 35 training questions, where ±1–3 points is within sampling noise. Please report Train and Test accuracies, per-split question counts, and confidence intervals (per-question bootstrap or per-seed variance) for every configuration.
  2. [Verified Acceptance (ε) and Effect of Candidate Pool Size N (Table VII, Figure 4)] The headline results are partly selected statistics. N=4 is chosen from the saturation curves in Figure 4/Table VII on SpreadsheetBench and ALFWorld—the same benchmarks that appear in the main results of Table I—and the regression threshold ε=0.05 is fixed with no sensitivity analysis. Because ε directly controls which candidates are committed, the authors should show how accuracy and Δ vary with ε (e.g., ε=0.0, 0.02, 0.05, 0.10), and should justify N=4 on a validation split or through a protocol that does not reuse the headline test benchmarks for hyperparameter selection.
  3. [Reproducibility (Experiments/Appendix)] No code, data, prompts, or generated skill files are released, and Table XI does not list the number of evolution iterations T used in Algorithm 1 (Figure 3 suggests five skill versions, but this is not stated). Together with the missing train-side numbers, this means the paper's main empirical claims cannot be independently reproduced or checked. I would consider this acceptable only if the authors release the skill-evolution harness, the exact prompts for the analyzer/generator, and the per-configuration train/test scores; otherwise the empirical contribution remains unverifiable.
minor comments (5)
  1. [Performance Comparison] The sentence 'per-model gains ranging from +10.6 to +28.4' is ambiguous: Table I's gains relative to No-Skill range from +2.3 to +47.4. Please specify the baseline and whether these are averages over benchmarks.
  2. [Theoretical Analysis] The symbol ε is used both for the regression threshold in Verified Acceptance (ε=0.05) and for the radius of the candidate ball in the theoretical analysis. These are different quantities and should use different symbols to avoid confusion.
  3. [Table XI] Table XI omits T, the number of evolution iterations. Please state the value used (Figure 3 implies five versions). Also clarify the relationship between the stated 'roughly 2.5× a single full evaluation' cost and the formula C(N)≈N|F|τ_fail + 2|D|τ_full.
  4. [Appendix: Statistical Analysis] The text refers to 'Table 1' instead of 'Table I.' The paired t-test and Wilcoxon test treat each model–benchmark pair as one observation; this is acknowledged, but the reader should be reminded that this does not capture per-question sampling variance, which is precisely the variance relevant to the Δ=Test−Train claim.
  5. [Figures and captions] Figure 1 has no labeled axes, and the caption says 'six skill versions' while Figure 3 shows five. Please align these details and label the axes.

Circularity Check

0 steps flagged

No significant circularity: SkillBoost's empirical claims are evaluated on held-out test splits; the acceptance gate and Theorem 1 are designed constraints and independent derivations, not fitted predictions.

full rationale

The central derivation chain is: roll out a frozen LLM on training split D, attribute failures, generate N candidate edits, and accept only candidates with r(s')>0 and Regress(s')<ε on D. Every optimization step is in-sample, but the reported SOTA results and the overfitting measure Δ=Test−Train are computed on held-out test splits (Tables I and III), so the outcome is not equivalent to the acceptance gate by construction. The gate is a deliberately chosen constraint with a fixed ε=0.05; no parameter is fitted to the test set and then renamed a prediction. Theorem 1 is derived from explicit sampling assumptions (uniform candidates in an ε-ball, Gaussian projection approximation) and is not fit to data; using it to explain N-saturation is illustrative rather than a circular prediction. The paper does not rely on load-bearing self-citations: baselines and related work are external, and no author-specific uniqueness theorem is invoked to force the design. The mildest concerns—selecting N=4 from the same benchmarks (Table VII) and not reporting the Train column or run-level variance in Table III—are selection-bias and verifiability issues, not circularity of the derivation. No quoted equation or fitted parameter reduces to its own input, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The ledger contains no new physical or metaphysical entities. The free parameters are design choices (ε,N,Top-K,T,guard samples) that the headline results depend on; N and ε are selected using the same benchmarks. The key domain assumptions are the decomposability of skills, the capability of the base LLM, and the proxy validity of the training split for generalization. The uniform-ball sampling assumption for Theorem 1 is explicitly ad hoc relative to the real candidate generator.

free parameters (5)
  • Regression threshold ε = 0.05
    Hand-set hyperparameter in Verified Acceptance (Fig. 2; Appendix Eq. 6). It defines what 'excessive regression' means and thus directly controls the overfitting claim; no sensitivity analysis is reported.
  • Candidate pool size N = 4 (default; ablations at 2/6/8)
    Selected from accuracy/cost ablation on the same benchmarks (Table VII, Fig. 4). Headline results use this tuned value.
  • Top-K cascaded selection count = 2
    Phase A keeps top-2 candidates for full-set evaluation (Appendix 'Two-Phase Cascaded Selection'); affects both cost and which edits get accepted.
  • Number of evolution iterations T = not reported (figures imply 5 rounds)
    Algorithm 1 takes T as input but it is missing from Table XI; results depend on when the evolution loop stops.
  • Guard-sample size = 10–20
    Random correct cases added to regression back-testing (Table XI); the anti-regression behavior depends on this sample count.
axioms (5)
  • domain assumption A skill is decomposable into independently editable components (Metadata, Workflow, Constraints, References), and edits to one component deterministically produce the next skill state.
    Section 'Self-Evolution Formulation' (state vector, deterministic transition P). If failures cannot be localized to a component, Structured Exploitation collapses to whole-skill editing.
  • domain assumption The frozen base LLM is capable enough to attribute failures, generate meaningful repair candidates, and evaluate candidates reliably.
    Used throughout Stages 1–3; the paper itself concedes in the Conclusion that 'weaker models may limit the quality of self-evolution.'
  • domain assumption Performance on the training split D plus the regression bound ε is a sufficient proxy for generalization to held-out tasks.
    Verified Acceptance commits edits based on D, and the overfitting claim is measured as Δ=Test−Train (Table III). If the training split is unrepresentative, the gate can still select overfit skills.
  • ad hoc to paper Theoretical candidates are i.i.d. uniform samples from an ϵ-ball around the current skill.
    Appendix Lemma 1 / Theorem 1. Actual candidates are generated by correlated LLM repair strategies, so the theorem is an analogy rather than a bound on the deployed algorithm.
  • domain assumption The five benchmark train splits are small enough to create genuine overfitting pressure and their test splits are representative of deployment.
    Experiment Setup; the generality of the overfitting-mitigation claim rests on this.

pith-pipeline@v1.3.0-daily-deepseek · 28301 in / 16677 out tokens · 169882 ms · 2026-08-01T11:40:26.348928+00:00 · methodology

0 comments
read the original abstract

Enabling large language model (LLM) agents to accumulate and reuse experience from past interactions remains a central challenge in real-world applications. A promising solution is to treat skills as trainable states and optimize them in the same way as model parameters in neural network training. However, data-driven skill optimization is prone to overfitting to the limited trajectories collected from real environments. Overexploiting these trajectories overfits the current batch, while unconstrained exploration causes regression on previously solved cases. This tension motivates a constrained search view of skill self-evolution, governed by an exploration--exploitation trade-off. We propose SkillBoost, a three-stage framework that mitigates both risks: structured exploitation localizes observed failures to editable skill components, prior-guided exploration draws on prior knowledge in the LLM to generate diverse repair candidates, and verified acceptance commits a candidate only when it improves performance within a regression bound. Experiments across 23 model--benchmark configurations show that SkillBoost achieves state-of-the-art performance while mitigating overfitting, outperforming both human-crafted and LLM-generated skills. Transfer experiments further show that optimized skills can be reused by other agents on similar tasks.

Figures

Figures reproduced from arXiv: 2607.26643 by Chao Liu, Hongqiang Lin, Nenggan Zheng, Xiaofan Bai, Xipeng Cao, Xuan Jin, Yuhong Li.

Figure 1
Figure 1. Figure 1: Effective skill updates should improve performance [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SkillBoost. Top: in each iteration, a frozen LLM agent executes tasks conditioned on the current skill st, and an acceptance gate commits only verified improvements to yield st+1. Bottom: backward optimization attributes failures to skill components (Structured Exploitation), generates N candidates under different repair strategies (Prior-guided Exploration), and selects the highest-gain candid… view at source ↗
Figure 4
Figure 4. Figure 4: Best-of-N results. Token Cost (bars, left axis) vs. Total Gain (line, right axis) as N grows. Dataset Final Acc. by N N=1 N=2 N=4 N=6 N=8 SpreadsheetBench 61.1 66.8 77.9 79.6 80.4 ALFWorld 74.9 76.8 82.0 82.3 83.1 TABLE VII: Best-of-N results. Accuracy improves as the candidate size N increases, but the gain saturates after N = 4. The shaded column denotes our setting. mization as an exploration–exploitati… view at source ↗
Figure 5
Figure 5. Figure 5: Repair Brief (Template): A structured diagnostic document with 8 modules for skill evolution. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: SkillBoost efficiency analysis across benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a) Ranking of skill line efficiency measured in percentage points gained per additional skill line, showing substantial [PITH_FULL_IMAGE:figures/full_fig_p017_7.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

103 extracted references · 22 linked inside Pith

  1. [1]

    Claude opus 4.5 system card, 2025

    Anthropic. Claude opus 4.5 system card, 2025

    2025

  2. [2]

    The scaling laws of skills in llm agent systems.arXiv preprint arXiv:2605.16508, 2026

    Charles Chen, Qiming Yu, Yuhang Gu, Zhuoye Huang, Hanjing Li, Hongyu Liu, Simin Liu, Jinhao Liu, Dengyun Peng, Jiangyi Wang, et al. The scaling laws of skills in llm agent systems.arXiv preprint arXiv:2605.16508, 2026

    Pith/arXiv arXiv 2026

  3. [3]

    A comprehensive survey of self- evolving AI agents: A new paradigm bridging foundation models and lifelong agentic systems.arXiv preprint arXiv:2508.07407, 2025

    Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, et al. A comprehensive survey of self- evolving AI agents: A new paradigm bridging foundation models and lifelong agentic systems.arXiv preprint arXiv:2508.07407, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Livemathematicianbench: A live benchmark for mathematician-level reasoning with proof sketches

    Linyang He, Qiyao Yu, Hanze Dong, Baohao Liao, Xinxing Xu, Micah Goldblum, Jiang Bian, and Nima Mesgarani. Livemathematicianbench: A live benchmark for mathematician-level reasoning with proof sketches. arXiv preprint arXiv:2604.01754, 2026

    arXiv 2026

  5. [5]

    Evotest: Evolutionary test-time learning for self-improving agen- tic systems

    Yufei He, Juncheng Liu, Yue Liu, Yibo Li, Tri Cao, Zhiyuan Hu, Xinxing Xu, and Bryan Hooi. Evotest: Evolutionary test-time learning for self-improving agen- tic systems. InInternational Conference on Learning Representations, 2026

    2026

  6. [6]

    Wei-Chieh Huang, Weizhi Zhang, Yueqing Liang, Yuanchen Bei, Yankai Chen, Tao Feng, Xinyu Pan, Zhen Tan, Yu Wang, Tianxin Wei, Shanglin Wu, Ruiyao Xu, Liangwei Yang, Rui Yang, Wooseong Yang, Chin- Yuan Yeh, Hanrong Zhang, Haozhen Zhang, Siqi Zhu, Henry Peng Zou, Wanjia Zhao, Song Wang, Wujiang Xu, Zixuan Ke, Zheng Hui, Dawei Li, Yaozu Wu, Langzhou He, Chen...

    arXiv 2026

  7. [7]

    From raw experience to skill consumption: A systematic study of model-generated agent skills.arXiv preprint arXiv:2605.23899, 2026

    Zisu Huang, Jingwen Xu, Yifan Yang, Ziyang Gong, Qihao Yang, Muzhao Tian, Xiaohua Wang, Changze Lv, Xuemei Gao, Qi Dai, et al. From raw experience to skill consumption: A systematic study of model-generated agent skills.arXiv preprint arXiv:2605.23899, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    SoK: Agentic skills–beyond tool use in llm agents.arXiv preprint arXiv:2602.20867, 2026

    Yanna Jiang, Delong Li, Haiyu Deng, Baihe Ma, Xu Wang, Qin Wang, and Guangsheng Yu. SoK: Agentic skills–beyond tool use in llm agents.arXiv preprint arXiv:2602.20867, 2026

    Pith/arXiv arXiv 2026

  9. [9]

    Organizing, orchestrating, and benchmarking agent skills at ecosys- tem scale.arXiv preprint arXiv:2603.02176, 2026

    Hao Li, Chunjiang Mu, Jianhao Chen, Siyue Ren, Zhiyao Cui, Yiqun Zhang, Lei Bai, and Shuyue Hu. Organizing, orchestrating, and benchmarking agent skills at ecosys- tem scale.arXiv preprint arXiv:2603.02176, 2026

    arXiv 2026

  10. [10]

    Muse-autoskill: Self-evolving agents via skill creation, memory, management, and evaluation.arXiv preprint arXiv:2605.27366, 2026

    Huawei Lin, Peng Li, Jie Song, Fuxin Jiang, and Tieying Zhang. Muse-autoskill: Self-evolving agents via skill creation, memory, management, and evaluation.arXiv preprint arXiv:2605.27366, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Skillforge: Forging domain-specific, self-evolving agent skills in cloud tech- nical support.arXiv preprint arXiv:2604.08618, 2026

    Xingyan Liu, Xiyue Luo, Linyu Li, Ganghong Huang, Jianfeng Liu, and Honglin Qiao. Skillforge: Forging domain-specific, self-evolving agent skills in cloud tech- nical support.arXiv preprint arXiv:2604.08618, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Spreadsheetbench: Towards challenging real world spreadsheet manipulation.Advances in Neural Informa- tion Processing Systems, 2024

    Zeyao Ma, Bohan Zhang, Jing Zhang, Jifan Yu, Xiaokang Zhang, Xiaohan Zhang, Sijia Luo, Xi Wang, and Jie Tang. Spreadsheetbench: Towards challenging real world spreadsheet manipulation.Advances in Neural Informa- tion Processing Systems, 2024

    2024

  13. [13]

    Skillclaw: Let skills evolve collectively with agentic evolver.arXiv preprint arXiv:2604.08377, 2026

    Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, and Xiangx- iang Chu. Skillclaw: Let skills evolve collectively with agentic evolver.arXiv preprint arXiv:2604.08377, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Minesh Mathew, Dimosthenis Karatzas, and C. V . Jawa- har. DocVQA: A Dataset for VQA on Document Images . InIEEE Winter Conference on Applications of Computer Vision, 2021

    2021

  15. [15]

    Trace2skill: Distill trajectory- local lessons into transferable agent skills.arXiv preprint arXiv:2603.25158, 2026

    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

    Pith/arXiv arXiv 2026

  16. [16]

    Skillos: Learning skill curation for self- evolving agents.arXiv preprint arXiv:2605.06614, 2026

    Siru Ouyang, Jun Yan, Yanfei Chen, Rujun Han, Zifeng Wang, Bhavana Dalvi Mishra, Rui Meng, Chun-Liang Li, Yizhu Jiao, Kaiwen Zha, Maohao Shen, Vishy Tiru- malashetty, George Lee, Jiawei Han, Tomas Pfister, and Chen-Yu Lee. Skillos: Learning skill curation for self- evolving agents.arXiv preprint arXiv:2605.06614, 2026

    Pith/arXiv arXiv 2026

  17. [17]

    The Berkeley function calling leaderboard (BFCL): From tool use to agentic evaluation of large lan- guage models

    Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji, Vishnu Suresh, Ion Stoica, and Joseph E Gonzalez. The Berkeley function calling leaderboard (BFCL): From tool use to agentic evaluation of large lan- guage models. InForty-second International Conference on Machine Learning, 2025

    2025

  18. [18]

    AutoRefine: From trajectories to reusable exper- tise for continual llm agent refinement.arXiv preprint arXiv:2601.22758, 2026

    Libin Qiu, Zhirong Gao, Junfu Chen, Yuhang Ye, Weizhi Huang, Xiaobo Xue, Wenkai Qiu, and Shuo Tang. AutoRefine: From trajectories to reusable exper- tise for continual llm agent refinement.arXiv preprint arXiv:2601.22758, 2026

    arXiv 2026

  19. [19]

    Qwen3.7-Plus: Multimodal agent intelli- gence, May 2026

    Qwen Team. Qwen3.7-Plus: Multimodal agent intelli- gence, May 2026

    2026

  20. [20]

    Qwen3.7: The agent frontier, May 2026

    Qwen Team. Qwen3.7: The agent frontier, May 2026

    2026

  21. [21]

    Skillopt: Trajectory- derived, verifier-grounded compilation of LLM-agent skills

    Rohan Rao and Vineeth Kalluru. Skillopt: Trajectory- derived, verifier-grounded compilation of LLM-agent skills. InFirst Workshop on Agent Skills, 2026

    2026

  22. [22]

    Skillopt-lite: Better and faster agent self-evolution via one line of vibe.arXiv preprint arXiv:2607.03451, 2026

    Yifei Shen, Bo Li, and Xinjie Zhang. Skillopt-lite: Better and faster agent self-evolution via one line of vibe.arXiv preprint arXiv:2607.03451, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Search and refine during think: Facilitating knowledge refinement for improved retrieval-augmented reasoning

    Yaorui Shi, Sihang Li, Chang Wu, Zhiyuan Liu, Jun- feng Fang, Hengxing Cai, An Zhang, and Xiang Wang. Search and refine during think: Facilitating knowledge refinement for improved retrieval-augmented reasoning. Advances in Neural Information Processing Systems, 2025

    2025

  24. [24]

    ALFWorld: Aligning text and embodied environments for interactive learning

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Cote, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. ALFWorld: Aligning text and embodied environments for interactive learning. InInternational Conference on Learning Representations, 2021

    2021

  25. [25]

    Reinforcement learning: An introduction 2nd ed.MIT press Cambridge, 2018

    Richard S Sutton, Andrew G Barto, et al. Reinforcement learning: An introduction 2nd ed.MIT press Cambridge, 2018

    2018

  26. [26]

    Skill-r1: Agent skill evolution via reinforcement learning.arXiv preprint arXiv:2605.09359, 2026

    Yash Vishe, Rohan Surana, Xunyi Jiang, Zihan Huang, Xintong Li, Nikki Lijing Kuang, Tong Yu, Ryan A Rossi, Jingbo Shang, Julian McAuley, et al. Skill-r1: Agent skill evolution via reinforcement learning.arXiv preprint arXiv:2605.09359, 2026

    Pith/arXiv arXiv 2026

  27. [27]

    SkillX: Automatically constructing skill knowledge bases for agents.arXiv preprint arXiv:2604.04804, 2026

    Chenxi Wang, Zhuoyun Yu, Xin Xie, Wuguannan Yao, Runnan Fang, Shuofei Qiao, Kexin Cao, Guozhou Zheng, Xiang Qi, Peng Zhang, and Shumin Deng. SkillX: Automatically constructing skill knowledge bases for agents.arXiv preprint arXiv:2604.04804, 2026

    Pith/arXiv arXiv 2026

  28. [28]

    V oyager: An open-ended embodied agent with large language models.Transactions on Machine Learning Research, 2024

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Man- dlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.Transactions on Machine Learning Research, 2024

    2024

  29. [29]

    Skillgrad: Optimizing agent skills like gradient descent.arXiv preprint arXiv:2605.27760, 2026

    Hanyu Wang, Yifan Lan, Bochuan Cao, Lu Lin, and Jinghui Chen. Skillgrad: Optimizing agent skills like gradient descent.arXiv preprint arXiv:2605.27760, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, Jun- yang Lin, Robert Brennan, Hao Peng, Heng Ji, and Graham Neubig. Openhands: An open platform for AI so...

    2025

  31. [31]

    Inducing programmatic skills for agentic tasks

    Zora Zhiruo Wang, Apurva Gandhi, Graham Neubig, and Daniel Fried. Inducing programmatic skills for agentic tasks. InSecond Conference on Language Modeling, 2025

    2025

  32. [32]

    Agent workflow memory

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Gra- ham Neubig. Agent workflow memory. InInternational Conference on Machine Learning, 2025

    2025

  33. [33]

    Chi, Chi Wang, Shuo Chen, Fernando Pereira, Wang-Cheng Kang, and Derek Zhiyuan Cheng

    Tianxin Wei, Noveen Sachdeva, Benjamin Coleman, Zhankui He, Yuanchen Bei, Xuying Ning, Mengting Ai, Yunzhe Li, Jingrui He, Ed H. Chi, Chi Wang, Shuo Chen, Fernando Pereira, Wang-Cheng Kang, and Derek Zhiyuan Cheng. Evo-memory: Benchmarking llm agent test-time learning with self-evolving memory. arXiv preprint arXiv:2511.20857, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    Skillrl: Evolving agents via recursive skill-augmented reinforcement learning.arXiv preprint arXiv:2602.08234, 2026

    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

    Pith/arXiv arXiv 2026

  35. [35]

    Agent skills for large language models: Architecture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430, 2026

    Renjun Xu and Yang Yan. Agent skills for large language models: Architecture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    SkillOpt: Executive strategy for self- evolving agent skills.arXiv preprint arXiv:2605.23904, 2026

    Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, et al. SkillOpt: Executive strategy for self- evolving agent skills.arXiv preprint arXiv:2605.23904, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Autoskill: Experience-driven lifelong learning via skill self-evolution.arXiv preprint arXiv:2603.01145, 2026

    Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, et al. Autoskill: Experience-driven lifelong learning via skill self-evolution.arXiv preprint arXiv:2603.01145, 2026

    arXiv 2026

  38. [38]

    Skilladaptor: Self-adapting skills for llm agents from trajectories

    Zhuoyun Yu, Xin Xie, Wuguannan Yao, Chenxi Wang, Lei Liang, Xiang Qi, and Shumin Deng. Skilladaptor: Self-adapting skills for llm agents from trajectories. arXiv preprint arXiv:2606.01311, 2026

    Pith/arXiv arXiv 2026

  39. [39]

    Co- EvoSkills: Self-evolving agent skills via co-evolutionary verification.arXiv preprint arXiv:2604.01687, 2026

    Hanrong Zhang, Shicheng Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayu Zhou, Chengze Li, Wei- Chieh Huang, Yifei Yao, Kening Zheng, et al. Co- EvoSkills: Self-evolving agent skills via co-evolutionary verification.arXiv preprint arXiv:2604.01687, 2026

    Pith/arXiv arXiv 2026

  40. [40]

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

    Boyuan Zheng, Michael Y . Fatemi, Xiaolong Jin, Zora Zhiruo Wang, Apurva Gandhi, Yueqi Song, Yu Gu, Jayanth Srinivasa, Gaowen Liu, Graham Neubig, and Yu Su. Skillweaver: Web agents can self-improve by discovering and honing skills.arXiv preprint arXiv:2504.07079, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Lifelongagentbench: Evaluating llm agents as lifelong learners.arXiv preprint arXiv:2505.11942, 2025

    Junhao Zheng, Xidi Cai, Qiuke Li, Duzhen Zhang, ZhongZhi Li, Yingying Zhang, Le Song, and Qianli Ma. Lifelongagentbench: Evaluating llm agents as lifelong learners.arXiv preprint arXiv:2505.11942, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    =SUM(A2:C2)

    Shanshan Zhong, Yi Lu, Jingjie Ning, Yibing Wan, Lihan Feng, Yuyi Ao, Leonardo FR Ribeiro, Markus Dreyer, Sean Ammirati, and Chenyan Xiong. Skill- learnbench: Benchmarking continual learning methods for agent skill generation on real-world tasks.arXiv preprint arXiv:2604.20087, 2026. APPENDIX To understand what the self-evolution process actually discover...

    Pith/arXiv arXiv 2026

  43. [43]

    Metadata Header Source version:v 1 (seed skill) T arget version:v 2 Model:Qwen-3.7-max Dataset:LiveMath train split (35 theorem-grounded multiple-choice questions, 5 options A–E) Baseline:v 1 accuracy = 22.9% (8/35), random baseline≈20%

  44. [44]

    Error analysis reveals two systematic failure modes induced by skill wording rather than model incapacity

    Baseline Performance Summary Per-category breakdown:nosep •Analysis/Proof questions: 2/12 (16.7%) •Computation questions: 4/15 (26.7%) •Conceptual questions: 2/8 (25.0%) Observation:Performance is near random chance across all categories. Error analysis reveals two systematic failure modes induced by skill wording rather than model incapacity

  45. [45]

    place only the option label inside the<answer>tag, no explana- tions

    Failure Mode Cluster Analysis Cluster A — Premature Answer without Reasoning. Root cause:Thev 1 output-format section instructs “place only the option label inside the<answer>tag, no explana- tions.” This cleanliness constraint was over-generalized by the model into “do not explain at all,” causing it to emit bare <answer>X</answer>responses (18 character...

  46. [46]

    reasoning zone

    Repair Strategy (Repair Actions) Action 1 (addressing Cluster A):Introduce a hard reasoning gate. Modify the output-format section to explicitly separate a “reasoning zone” (mandatory step-by-step analysis of each option) from an “answer zone” (the final<answer>tag). Add rule: “The<answer>tag must not appear until all options have been individually evalua...

  47. [47]

    Repair Action Mapping Table Failure Cluster Root Cause Repair Action SKILL.md Location A: Premature answer Reasoning gate absent Add mandatory reasoning zone§7 Output Format B: Strongest bias Over-strong selection rule Replace with exact-fit criterion§2 Selection Principle B: Strongest bias Missing over-strength check Add bidirectional verification§3 Veri...

  48. [48]

    •Output format compliance: all responses must contain exactly one<answer>tag

    Anti-Regression Guardrails Must-maintain strengths:nosep •Thev 1 skill’s 8 correctly solved cases must not regress. •Output format compliance: all responses must contain exactly one<answer>tag. Rejection criteria:nosep •If train accuracy drops belowv 1 baseline (22.9%)→reject mutation. •If regression count≥repaired cases→reject mutation. •If any response ...

  49. [49]

    •Cluster B (11 errors): 4 of 11 now apply bidirectional verification and select the exact-fit option

    Back-Testing Results (Post-evolution) T argeted re-test on previously failed cases:nosep •Cluster A (9 errors): all 9 now produce multi-paragraph reasoning; 2 of 9 reach the correct answer. •Cluster B (11 errors): 4 of 11 now apply bidirectional verification and select the exact-fit option. Comparison:nosep •Pre-repair (v 1): 8/35 correct (22.9%) •Post-re...

  50. [50]

    Nothing happens

    Execution Plan Step Status Collect failure trajectories fromv 1 evaluation✓ Cluster failures and identify root causes✓ Draft repair actions per cluster✓ Generatev 2 SKILL.md via guided editing✓ Back-testv 2 on full training set✓ Verify anti-regression guardrails✓ Accept or rejectv 2 ✓(accepted) This example illustrates how each module of the repair brief ...

  51. [55]

    Only choose admissible actions. ## 6. Common Errors - Blindly searching all containers and running out of steps. - Putting two objects into different containers or touching an already completed object. - Only examining instead of using the lamp. - Placing an object before the required transformation. - Ending before all goals are met. After evolution: ALF...

  52. [56]

    Break the task into locate, take, transform, and place

  53. [57]

    Search each place once, open containers before judging them empty, and prefer unvisited places

  54. [58]

    Take visible reachable targets immediately

  55. [59]

    Track how many objects remain

  56. [60]

    Only choose admissible actions. ## 6. Common Errors - Blindly searching all containers and running out of steps. - Reusing the first object in a two-object task instead of finding the second different object. - Taking a similar object with the wrong name. - Repeating the same action after Nothing happens. - Only examining instead of using the lamp. - Plac...

  57. [61]

    Initialization or authentication

  58. [62]

    Verification or analysis

  59. [63]

    ## Execution Strategy

    Summary or cleanup. ## Execution Strategy

  60. [64]

    Parse the request and identify all operations across all domains

  61. [65]

    Map each operation to the exact function

  62. [66]

    Plan the call sequence

  63. [67]

    Execute with no exploration, no repetition, and full domain coverage

  64. [68]

    ## Output Format Always respond with function calls when appropriate

    Verify that all expected functions were called. ## Output Format Always respond with function calls when appropriate. Do not explain or narrate. Just call the functions. After evolution: BFCLv 3.: --- skill_name: bfcl-solver current_version: v3 parent_version: v2 --- # BFCL Function Calling Solver ## Role You are an expert function calling agent. Given a ...

  65. [69]

    Parse the request to identify all required operations

  66. [73]

    ## Output Format Always respond with function calls when appropriate

    Verify that all expected functions have been called across all domains. ## Output Format Always respond with function calls when appropriate. Do not explain or narrate. Just call the functions. BFCL:v 3 →v 4: Before evolution: BFCLv 3.: --- skill_name: bfcl-solver current_version: v3 parent_version: v2 --- # BFCL Function Calling Solver ## Role You are an...

  67. [74]

    Parse the user request to identify all required operations

  68. [75]

    Map each operation to its specific function

  69. [76]

    Identify all API domains involved

  70. [77]

    Execute functions in logical order with no redundant calls

  71. [78]

    name": "function_name

    Verify that all expected functions have been called across all domains. ## Output Format Always respond with function calls when appropriate. Do not explain or narrate. Just call the functions. After evolution: BFCLv 4.: --- skill_name: bfcl-solver current_version: v4 parent_version: v3 --- # BFCL Function Calling Solver ## Role You are an expert function...

  72. [79]

    Am I using <function_calls> tags with valid JSON?

  73. [80]

    Are function names exact?

  74. [81]

    Have I included all functions from all conversation turns?

  75. [82]

    Are there any repeated calls?

  76. [83]

    Are there any unnecessary exploration or auth calls?

  77. [84]

    Definition Given a theorem-based multiple-choice math question, choose the single correct option

    Are all parameters correctly extracted? LiveMath:v 1 →v 2: Before evolution: LiveMathv 1.: --- type: task_skill task_name: LiveMath theorem MCQ solver current_version: v1 parent_version: earlier evolved version mutation_brief: repair direct guessing and over-strong choices --- # Skill: LiveMath Theorem-Grounded MCQ ## 1. Definition Given a theorem-based m...

  78. [86]

    A stronger or more complete statement often needs an extra assumption not given in the problem

    The correct option is not necessarily the strongest-looking statement. A stronger or more complete statement often needs an extra assumption not given in the problem. Such an option should be rejected.,→

  79. [88]

    When two options are close, compare them side by side and find the one wording difference. ## 3. Theorem-Level Precision First ask which option needs an extra assumption not stated in the problem. If an option needs extra smallness, extra regularity, stronger initial data, or a global assumption not given in the question, reject it.,→

  80. [89]

    Check over-strong options first

Showing first 80 references.