Pith.

Software Engineering · arXiv:2610.11858 v1 · submitted 8 Oct 2026 · 52 references

Trajectory-Guided Fault Localization for Agent Skill Evolution

Yu Ge, Linna Xie, Zhong Li, Yu Pei, Tian Zhang

The paper says

“Agent skills provide reusable guidance for code agents, but incomplete or unsuitable guidance can impair task execution. To reduce the manual effort of skill refinement, recent approaches use LLMs to generate revisions from execution feedback.”

more of the abstract

However, grounding these revisions in explicit behavioral evidence remains challenging. To address this gap, we propose SkillMorph, a skill-evolution approach based on trajectory-guided fault localization in agent skills. Its core idea is to link execution evidence to specific skill contents before generating revisions. Specifically, SkillMorph compares failure and success evidence in abstracted trajectories across repeated runs and tasks, incorporating changes between evolution loops to identify suspicious actions. It then uses these suspicious actions to localize edit sites in the skills and generate corresponding revisions. Experiments on SWE-Skills-Bench and CannBot show that the skills evolved by SkillMorph consistently achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods. We have also applied SkillMorph to automated kernel generation with an AI operator-development team, which has accepted 6 skill-revision pull requests.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.

Solid paper that adapts spectrum-based fault localization to agent skill evolution; the main risk is the unvalidated LLM-based step for classifying actions in successful runs, but the ablation and overall results still support the central claim. This line is wrong

3 major objections · 9 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 52 references say.

The paper says, at Section 4.2 (LLM-based error detection in successful runs)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The paper introduces a critical step where an LLM examines a window of h steps to decide whether an action in a successful run 'introduced an error state' (paragraph beginning 'If y indicates a success'). This binary decision determines whether the run provides failure or success coverage for that action.

more of objection 1

The paper reports no validation of this LLM judgment—no human evaluation, no agreement metrics (e.g., Cohen's kappa), no sensitivity analysis varying the window size h or the LLM provider. Since the entire suspiciousness scoring and thus the localization pipeline depends on these coverage counts, the reliability of this step is load-bearing. The authors should either (a) provide an empirical validation of this LLM-based detection (e.g., a human-annotated sample with agreement metrics), or (b) include a sensitivity analysis showing that downstream results are robust to plausible misclassification rates, or (c) explicitly discuss this limitation and why the overall approach can tolerate noise.

The paper says, at Section 4.2, historical adjustment (Eq. 6)

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The weight transformation w_{i,t}^σ(a) uses a specific functional form derived from Yule's Q. The paper states that the adjustment upweights increasing failure coverage and downweights decreasing failure coverage, but provides no justification for choosing this particular transformation over simpler alternatives (e.g., linear scaling, threshold-based clipping, or no adjustment).

more of objection 2

While the w/o History ablation in Table 2 shows that historical adjustment has a modest effect on overall performance, the specific choice of Eq. 6 could still affect the ranking of suspicious actions. The authors should either provide a rationale for the functional form or include an ablation comparing it to simpler alternatives. At minimum, the paper should note that this transformation is a heuristic and discuss whether results are sensitive to the exact formula.

The paper says, at Section 5.3, Table 1 and Figure 4 (regression analysis)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The paper reports that SkillMorph has fewer regressions than the baselines (9 vs 36 for CAP-Evolve on SWE overall, and 5 vs 11–19 skill instances). However, no statistical test is applied to assess whether these differences are significant given stochastic execution variance.

more of objection 3

The three-trial evaluation provides some robustness, but a paired bootstrap test or McNemar's test across the 98 skill instances would strengthen the claim that SkillMorph genuinely preserves existing behavior better. The authors should consider adding such a test or discussing the variability explicitly.

9 minor objections

The paper says, at Section 1, paragraph 2

Pith says: minor objection 1 of 9

The 'three challenges' are listed, but the third (deficiencies spanning multiple skills) overlaps with the first two in terms of motivation. Consider clarifying how the coordinated-revision aspect is a distinct challenge beyond the grounding and incidental-behavior challenges.

The paper says, at Section 4.1, trajectory abstraction

Pith says: minor objection 2 of 9

The paper cites prior work [11, 18] for the LLM-based trajectory abstraction. Reference [11] is from the same research group (FAMAS). The paper should clarify whether the abstraction method is used as-is or adapted, and if adapted, what changes were made.

show 7 more

The paper says, at Section 4.2, Eq. (5)

Pith says: minor objection 3 of 9

The ratio r_{i,t}^σ uses +1 in both numerator and denominator to handle zero counts. It would be helpful to note that this corresponds to Laplace (add-one) smoothing and to discuss the choice of the smoothing constant.

The paper says, at Section 4.3, edit-site localization

Pith says: minor objection 4 of 9

The description states that the LLM agent walks through the skill set 'recursively,' narrowing from skills to artifacts to contents. The paper should specify any limits on recursion depth or context length, and whether the process is fully automated or requires manual inspection at any step.

The paper says, at Table 1

Pith says: minor objection 5 of 9

The column headers 'SWE: A→B' and 'SWE: B→A' are clear, but the 'SWE: Overall' column could be more explicitly labeled (e.g., 'SWE: Overall (A+B)') to indicate it is the average of the two splits.

The paper says, at Figure 6

Pith says: minor objection 6 of 9

The y-axis label 'Success rate (%)' is consistent with the metrics but could be more precise, e.g., 'Pass@3 / Trial Acc. / Pass3 (%).'

The paper says, at References

Pith says: minor objection 7 of 9

Reference [6] (cap-evolve) links to a GitHub repository. If this is not a peer-reviewed publication, the paper should note that its description is based on the repository and specify the version or commit used, as implementations may change.

The paper says, at Section 6, Practical Application

Pith says: minor objection 8 of 9

The paper reports speedup statistics (64.26% achieve at least 0.6x, 54.14% at least 0.8x) for kernels generated with SkillMorph-evolved skills. It would be useful to compare these numbers against those obtained with the original skills or baseline-evolved skills to show that correctness improvements do not degrade performance.

The paper says, at General

Pith says: minor objection 9 of 9

Minor formatting: in the abstract and throughout, 'SkillMorph' sometimes appears with a leading space (e.g., ' SkillMorph'). Please ensure consistent formatting.

What would settle it

Construct a ground-truth set by manually annotating 50 successful trajectories from SWE-Skills-Bench (covering multiple skills and tasks). For each abstracted action, an expert labels whether the action introduced an error state from which the agent later recovered. Then run the LLM-based detector from Section 4.2 (same model, same h, same prompt) on the same trajectories and compute precision, recall, and F1 against the ground truth. If F1 < 0.8, the spectra are unreliable and the core localization evidence is suspect. Report the result; if poor, the central claim weakens.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Oct 2026

  1. Reader holds up confidence moderate
  2. Skeptic partly agrees: says it holds up only with conditions 31 s · 23,668 tokens read, 2,586 written LLM-based failure-detection in successful runs is unvalidated, weakening the reliability of suspicious-action spectra that drive the entire evolution pipeline.
  3. Editor wrote the verdict above 18 s · 24,532 tokens read, 1,584 written
  4. Referee major revision: 3 major objections, 9 minor 39 s · 24,019 tokens read, 3,628 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper is worth a serious look. It does something genuinely new: adapting spectrum-based fault localization (SBFL) to the problem of refining agent skills. Instead of dumping raw trajectories into an LLM and hoping for the best, SkillMorph first abstracts executions, identifies suspicious actions via Ochiai-like scores aggregated across runs, loops, and tasks, then traces those actions back to specific skill content via an LLM, and finally generates coordinated revisions. That pipeline is clearly motivated and well-structured.

The evaluation is the paper's strongest asset. Two diverse benchmarks (SWE-Skills-Bench, CannBot), held-out tasks, multiple trials, four baselines, and a careful ablation (Table 2). The results are consistent and practically meaningful: improved pass@3 and pass3, fewer regressions than any baseline, and actual accepted pull requests from a real deployment. The ablation shows that repeated execution matters most for coverage, and localization matters most for consistency.

Now the soft spots. The stress-test concern is real: the step that decides whether an action in a successful run "introduced an error state" (Section 4.2) is an unvalidated LLM judgment. No human evaluation, no agreement metrics, no sensitivity analysis. If that binary classification is systematically biased, the coverage counts that feed the suspiciousness scores are corrupted. The ablation (w/o Localization) shows localization helps overall, but it does not isolate this specific step. That said, the improvements are large enough that the pipeline is clearly doing something right even if the error detection is imperfect. I'd call this a moderate weakness, not a fatal one. The paper also has free parameters (number of runs, top-n actions, historical adjustment weights in Eq. 6) that are set without sensitivity analysis beyond the top-n sweep.

Who is this for? Researchers working on agent skill refinement, automated program repair for LLM agents, and practitioners building agent skill libraries. The paper is clearly written and the method is reproducible (code released).

Recommendation: Send to peer review. The central claim holds, the evaluation is strong, and the main weakness (unvalidated LLM judgment) can be addressed in revision with a targeted human evaluation or a synthetic perturbation analysis.

The strongest objection

The central claim depends on identifying suspicious actions via spectrum-based fault localization. A critical step (Section 4.2, paragraph beginning 'If y indicates a success') uses an LLM to decide whether an action in a successful run 'introduced an error state' by examining a window of h steps. This binary decision determines whether that run contributes failure or success coverage for the action. If the LLM systematically misclassifies (e.g., labeling recoverable errors as non-errors, or flagging harmless actions as error-introducing), the coverage counts N⁻ and N⁺ are corrupted, distorting the Ochiai-based suspiciousness scores and the subsequent ranking. The paper reports no validation of this LLM judgment—no human evaluation, no agreement metrics, no sensitivity analysis. Without such validation, the localization pipeline rests on an untrusted intermediate step. Even if cross-task aggregation and historical adjustment reduce noise, systematic bias could persistently misdirect revisions, or the observed improvements may come from the LLM's own revision abilities rather than the localization logic. The ablation in Table 2 (w/o Localization) shows localization improves pass3, but this does not validate the LLM’s judgment; it only shows that some form of localisation (possibly noisy) helps. The reliability of the central localisation mechanism remains unverified.

The reader's weakest_assumption concerns the premise that deficiencies manifest as recurring actions. That is a structural assumption about the problem domain. My concern targets the operational validity of the method that extracts the evidence for that recurrence: the LLM-based classification of actions in successful runs. If that classification is inaccurate, the entire suspicious-action spectrum—on which localization, historical adjustment, and cross-task aggregation depend—may be built on noisy or biased signals. The paper acknowledges no validation of this step, and ablation alone does not isolate its reliability. This is a load-bearing concern because erroneous coverage counts directly propagate to suspiciousness scores (Eq. 9) and thus to edit-site selection. The concrete test would settle whether this step is trustworthy; until then, the method's internal soundness is unverified. Hence I recommend CONDITIONAL: accept provided the authors validate or robustly analyze this LLM judgment.

The referee's report

The paper addresses a practically important problem—automating the refinement of agent skills used in LLM-based code agents. The proposed approach is novel in explicitly linking behavioral evidence from trajectories to skill content before generating revisions, drawing on the localize-then-repair paradigm from automated program repair. The evaluation is rigorous: it uses two benchmarks with different skill structures, held-out tasks, three independent trials, multiple baselines with official implementations, and ablation studies. If the results hold, SkillMorph could substantially reduce the manual effort of maintaining agent skill libraries and improve the reliability of code agents. The industrial case study (6 accepted pull requests) further demonstrates practical relevance.

Pith's reading of the paper (a machine's summary, not the authors')

SkillMorph is a method for automatically evolving the textual skills that guide LLM-based code agents. Its core insight is to treat agent actions as observable evidence of skill deficiencies, then adapt spectrum-based fault localization — comparing failure and success coverage across repeated runs and tasks — to identify suspicious actions. These actions are traced back to specific skill contents (edit sites), and coordinated revisions are generated across multiple skill artifacts. Experiments on two benchmarks show that skills evolved by SkillMorph improve trial-level accuracy by 15–100% relative to original skills and outperform four existing skill-evolution methods while introducing fewer regressions. The approach has also been deployed in an industrial AI-operator team, where six skill-revision pull requests were accepted.

What carries the argument. The central machinery is a spectrum-based fault localization adapted to agent trajectories: actions are treated as covered entities, and each run is treated as a test. Failure/success coverage counts are collected per action, adjusted by changes across evolution loops using a bounded transformation (inspired by Yule's Q), and aggregated across tasks. The suspiciousness score for an action is the product of a coverage ratio and a specificity ratio, following the Ochiai form. Suspicious actions then drive an LLM-based localization agent that recursively searches the skill set to identify the specific lines in SKILL.md or reference files that guided the problematic action, producing edit sites with a rationale.

Core claim. The paper claims that explicitly linking execution evidence to skill contents via trajectory-guided edit-site localization produces skills that are more effective and consistent than those from implicit feedback-driven revision. By abstracting trajectories into state transitions, comparing action coverage across runs and tasks with historical adjustment, and then localizing edit sites across multiple skills, SkillMorph generates revisions that address persistent deficiencies rather than incidental agent behavior. Empirical results on SWE-Skills-Bench and CannBot show consistent gains in pass@3, trial-level accuracy, and pass3 over original skills and four baselines, with larger improvements in execution consistency (pass3). On the CannBot benchmark, pass3 doubled from 46.67% to 93.33%.

The premise it rests on. The method assumes that skill deficiencies manifest as suspicious actions that recur across independent runs and tasks; if a deficiency only triggers sporadically or is masked by noise, the localization pipeline loses its foundation.

If the paper is right (Pith's inferences, not the authors')

If the central premise holds, skill evolution for LLM agents can be systematically grounded in execution evidence, reducing reliance on LLMs to implicitly diagnose problems from feedback.

The approach suggests that spectrum-based localization, a pillar of automated program repair, transfers to natural-language agent guidance when actions are used as covered entities.

Coordinated revisions across multiple skills become possible because a single suspicious action can be traced to edit sites in different skill artifacts, enabling holistic fixes.

The method's success on CannBot (a large, coordinated skill system) implies it scales to complex multi-file skill sets, not just independent single-file skills.

Because the method retains and reassesses historical evidence, later evolution loops can refine or revert earlier revisions, improving consistency without regressing previous gains.

An implicit consequence is that the quality of skill evolution depends critically on the LLM's ability to correctly identify edit sites from suspicious actions. If the LLM mislocalizes, the revision may be ineffective; the paper acknowledges this threat but does not measure localization accuracy separately.

The method may generalize beyond software engineering to any domain where agents follow natural-language guidance and task outcomes are measurable, such as data analysis or document processing.

A testable extension: applying SkillMorph to a benchmark where skill deficiencies are artificially injected (e.g., by removing a crucial instruction) would reveal whether the method reliably localizes the injected fault across varying amounts of stochastic agent behavior.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Agent skills provide reusable guidance for code agents, but incomplete or unsuitable guidance can impair task execution. To reduce the manual effort of skill refinement, recent approaches use LLMs to generate revisions from execution feedback. However, grounding these revisions in explicit behavioral evidence remains challenging. To address this gap, we propose SkillMorph, a skill-evolution approach based on trajectory-guided fault localization in agent skills. Its core idea is to link execution evidence to specific skill contents before generating revisions. Specifically, SkillMorph compares failure and success evidence in abstracted trajectories across repeated runs and tasks, incorporating changes between evolution loops to identify suspicious actions. It then uses these suspicious actions to localize edit sites in the skills and generate corresponding revisions. Experiments on SWE-Skills-Bench and CannBot show that the skills evolved by SkillMorph consistently achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods. We have also applied SkillMorph to automated kernel generation with an AI operator-development team, which has accepted 6 skill-revision pull requests.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-09

The authors' main claim, as the classifier picked it out 1.00

“Experiments on SWE-Skills-Bench and CannBot show that the skills evolved by SkillMorph consistently achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.10 · human participants 0.27 · controlled comparison 0.89 · randomized assignment 0.02 · states quantitative result 0.86 · compares with prior results 0.98 · mentions code or data availability 0.05

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Agent skills provide reusable guidance for code agents, but incomplete or unsuitable guidance can impair task execution. background 1.00 0.05 0.02 asserted 0.79
To reduce the manual effort of skill refinement, recent approaches use LLMs to generate revisions from execution feedback. background 1.00 0.04 0.02 asserted 1.00
However, grounding these revisions in explicit behavioral evidence remains challenging. gap 0.99 0.05 0.02 asserted 0.98
To address this gap, we propose SkillMorph, a skill-evolution approach based on trajectory-guided fault localization in agent skills. contribution 1.00 0.83 0.05 asserted 0.97
Its core idea is to link execution evidence to specific skill contents before generating revisions. method 0.73 0.52 0.04 asserted 1.00
Specifically, SkillMorph compares failure and success evidence in abstracted trajectories across repeated runs and tasks, incorporating changes between evolution loops to identify suspicious actions. method 1.00 0.87 0.04 asserted 0.98
It then uses these suspicious actions to localize edit sites in the skills and generate corresponding revisions. method 1.00 0.81 0.04 asserted 1.00
Experiments on SWE-Skills-Bench and CannBot show that the skills evolved by SkillMorph consistently achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods. result 1.00 0.96 0.80 asserted 1.00
We have also applied SkillMorph to automated kernel generation with an AI operator-development team, which has accepted 6 skill-revision pull requests. result 0.90 0.95 0.06 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-09: 1 request, 6,792 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (6)
Figure 1 of the paper
Figure 1. Structure of the mcp-builder skill. In this work, we focus on code agents [34, 43] that use skills for software engineering tasks and refer to the col￾lection of skills available to an agent as its skill set. A skill set provides procedural knowledge and project-specific context for a class of related tasks without modifying the underlying model parameters [19, 42]. Following the Agent Skills specification [2], each skill is a file-based package with a SKILL.md document as its entry point, as illustrated in
Figure 2 of the paper
Figure 2. An example of SkillMorph evolving the mcp-builder skill. For illustration, we use document query (Task 1) and note management (Task 2) for evolution, and hold out Task 3, which indexes Markdown documents for search and retrieval. Failures under the original skill
Figure 3 of the paper
Figure 3. Overview of SkillMorph. tion evidence, analyses, and revision history to support these analyses. After L loops, SkillMorph returns the resulting skill set SL. 4.1 Execution and Trajectory Abstraction This stage collects execution evidence under the current skill set to support subsequent edit-site lo￾calization and revision generation. SkillMorph first executes each evolution task multiple times and retains all runs, regardless of their outcomes. It then abstracts each raw trajectory into a uni￾form state-transition representation that captures actions and their observed effects,…
Figure 4 of the paper
Figure 4. Fine-grained performance changes relative to the original skills.
Figure 5 of the paper
Figure 5. Effect of the Number of Evolution Loops on
Figure 6 of the paper
Figure 6. Effect of top-n action selection. and vary n among 1, 5, 10, 15, and 20. All other experimental settings remain unchanged

Figures are the authors', reproduced from the paper.

References (52)
  1. [1] Salaheddin Alzubi, Noah Provenzano, Jaydon Bingham, Weiyuan Chen, and Tu Vu. Evoskill: Automated skill discovery for multi-agent systems.CoRR, abs/2603.02766, 2026 on PithPith's reviewarXiv
  2. [2] Anthropic. Agent Skills specification, 2025 on Pith
  3. [3] Anthropic. Claude Code: Overview. Online documentation, 2026 on Pith
  4. [4] Ascend. Npukernelbench, 2026 on Pith
  5. [5] Islem Bouzenia and Michael Pradel. Understanding software engineering agents: A study of thought-action-result trajectories. In40th IEEE/ACM International Conference on Automated 16 Software Engineering, ASE 2025, Seoul, Korea, Republic of, November 16-20, 2025, pages 2846– on Pith
  6. [6] cap-evolve: a skills-native, host-agnostic harness for honestly optimizing ai-agent capabilities, on Pith
  7. [7] Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya G. Parameswaran, Dan Klein, Kannan Ramchandran, Matei A. Zaharia, Joseph E. Gonzalez, and Ion Stoica. Why do multi-agent LLM systems fail? In Danielle Belgrave, Chen... on Pith
  8. [8] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pond´ e de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott... on PithPith's reviewarXiv
  9. [9] DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence.CoRR, abs/2606.19348, 2026 on PitharXiv
  10. [10] Srishti Gautam, Arjun Radhakrishna, and Sumit Gulwani. Skillaxe: Sharpening llm-authored agent skills through evaluation-guided self-refinement.CoRR, abs/2606.10546, 2026 on PithPith's reviewarXiv
  11. [11] Yu Ge, Linna Xie, Zhong Li, Yu Pei, and Tian Zhang. Spectrum-based failure attribution for multi-agent systems.Proc. ACM Softw. Eng., 3(FSE), June 2026 on Pith
  12. [12] Tingxu Han, Yi Zhang, Wei Song, Chunrong Fang, Zhenyu Chen, Youcheng Sun, and Lijie Hu. Swe-skills-bench: Do agent skills actually help in real-world software engineering?CoRR, abs/2603.15401, 2026 on PithPith's reviewarXiv
  13. [13] Mark Harman and Bryan F. Jones. Search-based software engineering.Information and Software Technology, 43(14):833–839, 2001 on Pith
  14. [14] David Boram Hong, Aaron Imani, and Iftekhar Ahmed. From anatomy to smells: An empirical study of skill.md in agent skills, 2026 on Pith
  15. [15] Wei-Jung Huang. Do LLM-authored agent skills help? a component-level ablation in data analysis tasks. OpenReview, 2026 on Pith
  16. [16] Jeffrey O. Kephart and David M. Chess. The vision of autonomic computing.Computer, 36(1):41– 50, 2003 on Pith
  17. [17] Gaoyuan Li, Meihao Fan, Yizhe Liu, Shaolei Zhang, Ju Fan, Siyi Wang, Jiaheng Hou, Xudong Weng, Honghan Tian, and Zang Li. Skilladam: Stable and efficient skill evolution for agents. arXiv preprint arXiv:2609.08944, 2026 on PithPith's reviewarXiv
  18. [18] Han Li, Yifan Yao, Letian Zhu, Rili Feng, Hongyi Ye, Jiaming Wang, Yancheng He, Pengyu Zou, Lehan Zhang, Xinping Lei, Haoyang Huang, Ken Deng, Ming Sun, Zhaoxiang Zhang, He Ye, and Jiaheng Liu. Codetracer: Towards traceable agent states.CoRR, abs/2604.11641, 2026. 17 on PithPith's reviewarXiv
  19. [19] Xiangyi Li, Wenbo Chen, Yimin Liu, Shenghan Zheng, Xiaokun Chen, Yifeng He, Yubo Li, Bingran You, Haotian Shen, Jiankai Sun, Shuyi Wang, Qunhong Zeng, Di Wang, Xuandong Zhao, Yuanli Wang, Roey Ben Chaim, Zonglin Di, Yipeng Gao, Junwei He, Yizhuo He, Liqiang Jing, Luyang Kong, ... on PithPith's reviewarXiv
  20. [20] Heng Liao, Jiajin Tu, Jing Xia, Hu Liu, Xiping Zhou, Honghui Yuan, and Yuxing Hu. Ascend: a scalable and unified architecture for ubiquitous deep neural network computing : Industry track paper. InIEEE International Symposium on High-Performance Computer Architecture, HPCA 202... on Pith
show 32 more
  1. [21] Shukai Liu, Bo Jiang, Jian Yang, Yizhi Li, Jinyang Guo, Xianglong Liu, and Bryan Dai. Context as a tool: Context management for long-horizon swe-agents. In Maria Liakata, Viviane P. Mor- eira, Jiajun Zhang, and David Jurgens, editors,Findings of the Association for Computation... on Pith
  2. [22] Shuyang Liu, Yang Chen, Rahul Krishna, Saurabh Sinha, Jatin Ganhotra, and Reyhaneh Jabbar- vand. Process-centric analysis of agentic software systems.Proc. ACM Program. Lang., 10(OOP- SLA1), April 2026 on Pith
  3. [23] Yuxuan Liu, Zhaochen Su, Lingyun Xie, Yuhao Zhang, Qing Zong, Jiahe Guo, Zhongwei Xie, Yiyan Ji, Yauwai Yim, Hongyu Luo, Xiyu Ren, Ruan Chenyu, Haoran Li, and Yangqiu Song. Skillrevise: Improving llm-authored agent skills via trace-conditioned skill revision.CoRR, abs/2606.01139, 2026 on PithPith's reviewarXiv
  4. [24] Zijian Lu, Yiping Zuo, Yupeng Nie, Xin He, Weibei Fan, Lianyong Qi, and Shi Jin. Contractskill: Repairable contract-based skills for multimodal web agents.CoRR, abs/2603.20340, 2026 on PithPith's reviewarXiv
  5. [25] Oorja Majgaonkar, Zhiwei Fei, Xiang Li, Federica Sarro, and He Ye. Understanding code agent behaviour: An empirical study of success and failure trajectories.CoRR, abs/2511.00197, 2025 on PitharXiv
  6. [26] 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.CoRR, abs/2603.25158, 2026 on PithPith's reviewarXiv
  7. [27] Akira Ochiai. Zoogeographical studies on the soleoid fishes found in japan and its neighbouring regions-i.Nippon Suisan Gakkaishi, 22(9):522–525, 1957 on Pith
  8. [28] Jincheng Ren, Siwei Wu, Yizhi Li, Kang Zhu, Shu Xu, Boyu Feng, Ruibin Yuan, Wei Zhang, Riza Batista-Navarro, Jian Yang, and Chenghua Lin. A self-evolving framework for efficient terminal agents via observational context compression.CoRR, abs/2604.19572, 2026 on PithPith's reviewarXiv
  9. [29] Benjamin Rombaut, Sogol Masoumzadeh, Kirill Vasilevski, Dayi Lin, and Ahmed E. Hassan. Watson: A cognitive observability framework for the reasoning of llm-powered agents. In40th IEEE/ACM International Conference on Automated Software Engineering, ASE 2025, Seoul, Korea, Repub... on Pith
  10. [30] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflex- ion: language agents with verbal reinforcement learning. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors,Advances in Neural Informa- t... on Pith
  11. [31] Yu Tian, Jiawei Chen, Lifan Zheng, Mingxiang Tao, Xinyi Zeng, Zhaoxia Yin, Hang Su, and Xian Sun. Skills-coach: A self-evolving skill optimizer via training-free GRPO.CoRR, abs/2604.27488, 2026. 18 on PithPith's reviewarXiv
  12. [32] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Trans. Mach. Learn. Res., 2024, 2024 on Pith
  13. [33] Hanyu Wang, Yifan Lan, Bochuan Cao, Lu Lin, and Jinghui Chen. Skillgrad: Optimizing agent skills like gradient descent.CoRR, abs/2605.27760, 2026 on PithPith's reviewarXiv
  14. [34] 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, Junyang Lin, et al. Openha... on Pith
  15. [35] Matthijs J. Warrens. On association coefficients for 2×2 tables and properties that do not depend on the marginal distributions.Psychometrika, 73(4):777–789, 2008 on Pith
  16. [36] Westley Weimer, ThanhVu Nguyen, Claire Le Goues, and Stephanie Forrest. Automatically find- ing patches using genetic programming. In31st International Conference on Software Engineering, ICSE 2009, May 16-24, 2009, Vancouver, Canada, Proceedings, pages 364–374. IEEE, 2009 on Pith
  17. [37] Ming Wen, Junjie Chen, Rongxin Wu, Dan Hao, and Shing-Chi Cheung. Context-aware patch generation for better automated program repair. In Michel Chaudron, Ivica Crnkovic, Marsha Chechik, and Mark Harman, editors,Proceedings of the 40th International Conference on Software Engin... on Pith
  18. [38] W. Eric Wong, Ruizhi Gao, Yihao Li, Rui Abreu, and Franz Wotawa. A survey on software fault localization.IEEE Trans. Software Eng., 42(8):707–740, 2016 on Pith
  19. [39] Yuan-An Xiao, Pengfei Gao, Chao Peng, and Yingfei Xiong. Reducing cost of llm agents with trajectory reduction.Proc. ACM Softw. Eng., 3(FSE), June 2026 on Pith
  20. [40] Linna Xie, Zhong Li, Yu Pei, Zhongzhen Wen, Kui Liu, Tian Zhang, and Xuandong Li. Premm: Llm-based program repair for multi-method bugs via divide and conquer.Proc. ACM Program. Lang., 9(OOPSLA2):1316–1344, 2025 on Pith
  21. [41] Yingfei Xiong, Xinyuan Liu, Muhan Zeng, Lu Zhang, and Gang Huang. Identifying patch cor- rectness in test-based program repair. In Michel Chaudron, Ivica Crnkovic, Marsha Chechik, and Mark Harman, editors,Proceedings of the 40th International Conference on Software Engineering... on Pith
  22. [42] Renjun Xu and Yang Yan. Agent skills for large language models: Architecture, acquisition, security, and the path forward.CoRR, abs/2602.12430, 2026 on PithPith's reviewarXiv
  23. [43] John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. ... on Pith
  24. [44] Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao, Qi Dai, Bei Liu, Kai Qiu, Yuqing Yang, Dongdong Chen, Xue Yang, and Chong Luo. Skillopt: Executive strategy for self-evolving agent skills.CoRR, abs/2605.23904, 2026 on PithPith's reviewarXiv
  25. [45] Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik R Narasimhan.τ-bench: A benchmark for Tool-Agent-User interaction in real-world domains. InThe Thirteenth International Conference on Learning Representations, 2025 on Pith
  26. [46] Zhuoyun Yu, Xin Xie, Wuguannan Yao, Chenxi Wang, Lei Liang, Xiang Qi, and Shumin Deng. Skilladaptor: Self-adapting skills for LLM agents from trajectories.CoRR, abs/2606.01311, 2026. 19 on PithPith's reviewarXiv
  27. [47] Barry Zhang, Keith Lazuka, and Mahesh Murag. Equipping agents for the real world with Agent Skills. Anthropic Engineering, October 2025 on Pith
  28. [48] Genrui Zhang, Erle Zhu, Jinfeng Zhou, Caiyan Jia, and Hongning Wang. Skillevolver: Skill learning as a meta-skill.CoRR, abs/2605.10500, 2026 on PithPith's reviewarXiv
  29. [49] Hanrong Zhang, Shicheng Fan, Henry Peng Zou, Yankai Chen, Zhenting Wang, Jiayu Zhou, Chengze Li, Wei-Chieh Huang, Yifei Yao, Kening Zheng, Xue Liu, Xiaoxiao Li, and Philip S. Yu. Coevoskills: Self-evolving agent skills via co-evolutionary verification.CoRR, abs/2604.01687, 2026 on PithPith's reviewarXiv
  30. [50] Shaokun Zhang, Ming Yin, Jieyu Zhang, Jiale Liu, Zhiguang Han, Jingyang Zhang, Beibin Li, Chi Wang, Huazheng Wang, Yiran Chen, and Qingyun Wu. Which agent causes task failures and when? on automated failure attribution of LLM multi-agent systems. In Aarti Singh, Maryam Fazel, ... on Pith
  31. [51] Xiangxin Zhao, Han Li, Shuaiting Li, Tianyi Zhao, Earl T. Barr, Federica Sarro, and He Ye. Failure as a process: An anatomy of cli coding agent trajectories, 2026. 20 on Pith
  32. [2026] https://github.com/skillberry-ai/cap-evolve on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment