Software Engineering · arXiv:2608.06153 v1 · submitted 6 Aug 2026 · 37 references
Learning Globally Reusable Skills for Coding Agents
The paper says
“Automated skill evolution enables Large Language Model (LLM) agents to continuously improve without expensive retraining.”
more of the abstract
However, existing approaches typically treat skill evolution as a sequence of local updates, overlooking relationships among skills and often producing overfitted skill updates that fail to generalize across tasks. We propose GSE, a globalized skill evolution framework that jointly optimizes skill compatibility and skill generalization. To preserve consistency across the skill bank, GSE maintains a Skill Relation Graph (SRG) that explicitly models and co-evolves inter-skill relationships. To improve generalization, GSE performs cluster-based skill consolidation to abstract reusable capabilities from local updates and employs replay-driven verification to prevent overfitting and behavioral regressions. We evaluate GSE on two representative software engineering tasks: bug-revealing test generation and false-positive bug report filtering. Across two state-of-the-art coding agents, OpenHands and mini-SWE-agent, GSE consistently achieves the best precision, recall, and F1-score. Compared with existing evolution techniques, GSE improves precision and recall by 6.1% 34.1% and 31.8% 180.0% for test generation, and by 15.4% 96.4% and 13.1% 19.8% for false-positive filtering. Deployment on an internal industrial agent further yields a 61.4% improvement in F1-score, demonstrating the effectiveness and generalizability of GSE for evolving effective skills.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 7 Aug 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A genuinely new skill-evolution framework with a solid held-out evaluation, but the headline numbers need error bars and a fixed comparison set. This line is wrong
5 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 37 references say.
The paper says, at §5.1.2, Table 1; §5.2.2, Table 2
The review did not quote this passage.
Pith says: Objection 1 of 5 · major
The central claim that GSE 'consistently' achieves the best precision, recall, and F1 is supported only by pooled point estimates from a single evaluation pass. For test generation there are only 9 held-out projects (108 bugs), and for filtering only 8 repositories (500 reports).
more of objection 1
The smallest claimed gain, OpenHands precision 0.35 versus 0.33 (6.1% relative), could be reversed by a small number of reclassified cases. No per-project breakdown, no confidence intervals, and no paired significance test are reported. Because the one-project-held-out design yields paired observations (each method is evaluated on the same held-out project), a paired test such as the Wilcoxon signed-rank test over projects, or repeated runs with different seeds, is feasible and should be reported before the abstract's 'consistently' is justified.
The paper says, at Abstract; §5.1.2, §9
The review did not quote this passage.
Pith says: Objection 2 of 5 · major
The abstract and conclusion state that 'compared with existing evolution techniques,' GSE improves test-generation recall by 31.8% to 180.0%. The 180.0% figure is not computed against an evolution technique: it is the improvement over the Human Skill baseline on OpenHands (recall 0.28 versus 0.10). The largest improvement over an actual evolution technique (Live-SWE-agent or Trace2Skill) is 64.7%. The comparison set should be stated consistently in the abstract, the results section, and the conclusion.
The paper says, at §3.3, Algorithm 1
The review did not quote this passage.
Pith says: Objection 3 of 5 · major
The skill generalization mechanism depends on several free parameters that are never operationally defined: the replay acceptance criterion ('preserves or improves performance,' line 22), the semantic-similarity threshold for clustering new-skill proposals (line 12), and the trace-simplification filters in §3.1. These choices can materially affect which skills are accepted and, consequently, the reported gains. Without specifying these thresholds or releasing code and configuration, the method is not fully reproducible, and the reader cannot assess sensitivity of the results to these choices.
The paper says, at §4.1.1
The review did not quote this passage.
Pith says: Objection 4 of 5 · major
The test-generation benchmark is manually filtered: the authors state that they 'manually verify whether invoking the method can trigger the corresponding bug, and retain only verified triggerable methods.' No inter-rater reliability, rater instructions, or release of the filtered benchmark is provided.
more of objection 4
Since the entire test-generation evaluation rests on this filtered set, the authors should either make the filtered benchmark available with labels, or report a secondary analysis on the unfiltered Multi-SWE-Bench Java subset to demonstrate that the manual filtering does not drive the results.
The paper says, at §3.2.2
The review did not quote this passage.
Pith says: Objection 5 of 5 · major
The Skill Relation Graph is central to GSE, but the paper does not describe how initial graph edges are constructed, how the update agent reliably infers dependency, co-usage, or conflict edges from traces, or how erroneous edges are corrected. Replay-driven validation only checks the retrieved historical cases, so an incorrect edge that is not exercised by those cases can persist and propagate incorrect auxiliary proposals to unrelated skills.
more of objection 5
A sensitivity analysis or a small error analysis of SRG edge quality would substantially strengthen the attribution of the observed gains to relationship-aware evolution.
5 minor objections
The paper says, at §4.3
Pith says: minor objection 1 of 5
The Trace2Skill bullet ends with a comma and an incomplete clause ('merges them at once into one skill bank,'); this should be completed and ended with a period.
The paper says, at §3.2.3
Pith says: minor objection 2 of 5
The reference to 'Figure 3.2.2' should be 'Figure 2'.
show 3 more
The paper says, at §6, Table 5
Pith says: minor objection 3 of 5
The industrial deployment compares only the base agent and human-written skills with GSE; reporting the evolution baselines (Live-SWE-agent and Trace2Skill) on the internal agent would make the deployment results directly comparable with Tables 1 and 2.
The paper says, at §7
Pith says: minor objection 4 of 5
The statement that 'None of the generated tests align with the reference tests' should specify the alignment check (e.g., exact string match versus structural equality) and report the number of cases checked.
The paper says, at §5.4, Table 4
Pith says: minor objection 5 of 5
Table 4 reports token costs macro-averaged across tasks and agents; the text in §5.4.2 explains the exclusion of Live-SWE-agent for evolution cost, but the table caption should restate this caveat so the reader does not misinterpret the comparison.
What would settle it
Request or reconstruct per-project held-out results for GSE and the strongest baseline (at least Trace2Skill, ideally Live-SWE-agent too) for both agents and tasks; then compute paired bootstrap 95% confidence intervals over the 9 (or 8) held-out projects for precision, recall, and F1, and count how many projects favor GSE per metric. If the confidence interval for any claimed gain (especially the 6.1% precision and 13.1% recall differences) includes zero, or if fewer than 7 of 9 projects favor GSE on that metric, the "consistently" claim is not supported and the verdict should require additional runs or released per-project artifacts.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 7 Aug 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 5 major objections, 5 minor
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
What you should know: GSE is a real step forward in skill evolution for LLM coding agents. The core idea is to treat the skill bank as a graph with dependency, co-usage and conflict edges, co-evolve those relations alongside skill content, then consolidate local proposals by clustering and verify by replay. That is genuinely different from Trace2Skill and Live-SWE-agent, which merge updates locally. The paper is also honest about the forward-test setup: held-out projects are never used to evolve the bank, so the protocol is a real generalization test.
What it does well: the evaluation design is solid for a systems paper. Two base agents, two tasks, an ablation that isolates the graph and the generalization mechanism, and an industrial deployment. The numbers are internally consistent and the improvements are large in most cells. The case study in Figure 3 illustrates the failure mode of local updates clearly.
Where the soft spots are: the statistics are thin. All results are pooled point estimates, with no confidence intervals, no per-project breakdown, and no significance tests. With 9 held-out projects, the smallest claimed gain—6.1% precision on OpenHands—could easily be a single project's noise. The abstract says "compared with existing evolution techniques" but the 180% recall improvement is relative to human-written skills, not an evolution baseline; the comparison set shifts between cells. That should be fixed. Also, no code or data is released, the industrial dataset is proprietary, and CodeSkill/Skill1 are cited in related work but not included as baselines. Those omissions limit independent verification.
None of this kills the main contribution. The relation graph plus consolidation-plus-replay is a plausible mechanism for avoiding overfit, and the held-out evaluation is a genuine forward test, not a circular one. But the paper needs a revision with error bars or per-project results, a fixed comparison set, and at least one external baseline before the claims are load-bearing.
Who this is for: anyone working on agent skill banks or self-evolving coding agents. It deserves a serious referee. I'd recommend acceptance with major revision expectations rather than desk reject.
Recommendation: send it out, but insist the authors add per-project results, paired significance tests, and a precise statement of what each improvement percentage is relative to.
The strongest objection
The central claim is empirical: GSE "consistently" achieves the best precision, recall, and F1 on held-out projects. The evidence in Tables 1-2 consists of one pooled aggregate per agent-task, with no per-project results, no confidence intervals, and no repeated runs. The evaluation uses only 9 Java projects (108 bugs) for test generation and 8 industrial repos (500 reports) for filtering. Under a one-project-held-out protocol, a single project is withheld at a time, so the smallest reported advantage—OpenHands precision 0.35 vs. 0.33 (+0.02, 6.1%)—can be shifted by one or two test cases changing outcome. Setting temperature to zero does not eliminate variability in multi-step LLM agent trajectories or test execution. The abstract's 180% recall improvement is computed against human-written skills rather than an evolution technique, which suggests the "compared with existing evolution techniques" range is not a fixed comparison set. Without a per-project breakdown and a paired significance test over projects, we cannot tell whether GSE is systematically better or whether the aggregate gap is driven by a single favorable held-out run.
I read GSE charitably: the Skill Relation Graph and cluster-based consolidation are coherent contributions, the ablation is directionally consistent with the design, and the one-project-held-out protocol is a genuine attempt to test cross-project generalization. The industrial deployment, though uncontrolled, adds some ecological validity. The load-bearing assertion, however, is the empirical headline: GSE "consistently" achieves the best precision, recall, and F1 on held-out projects. That claim rests on aggregate point estimates from a small number of held-out projects. With 108 Java bugs across 9 projects, the test-generation precision difference between GSE and the strongest baseline on OpenHands is 0.02; a single project's held-out run can dominate the aggregate. The paper provides no per-project breakdown, no error bars, and no significance tests. Temperature zero does not guarantee reproducibility in agentic pipelines, where LLM API nondeterminism, retrieval order, and tool execution can still vary. The reader's weakest_assumption focuses on the correctness of the LLM-generated Skill Relation Graph and whether replay is a sufficient proxy for generalization; my concern is one level upstream. Even if the graph and replay behave exactly as designed, the reported margin has not yet been shown to be outside noise. This is a correctness-risk concern about the evaluation, not a novelty or consensus objection. The concrete check—per-project bootstrap over held-out projects—would settle the concern without requiring a full reimplementation. It is a reasonable condition for accepting the central claim as verified, which is why I keep the reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT.
The referee's report
If the empirical results hold, GSE would be a meaningful contribution to automated skill evolution for coding agents. The evaluation design has notable strengths: a forward one-project-held-out protocol, two independent base agents, two complementary tasks, ablations of the two main components, and an industrial deployment. The paper is also careful to check for test leakage against reference tests and to control the base LLM and execution environment. However, the central quantitative claim of consistent superiority is currently supported only by pooled point estimates without statistical grounding, and the method description leaves several load-bearing parameters unspecified. These issues are addressable and do not undermine the overall approach, but they must be resolved before the claim of consistent improvement can be accepted.
Pith's reading of the paper (a machine's summary, not the authors')
Automated skill evolution lets coding agents improve by updating a bank of reusable procedural knowledge instead of retraining the model. Existing methods treat each update as an isolated local patch, which can overfit to one execution trace and clash with other skills. GSE claims that skill evolution should be a global optimization: it maintains a Skill Relation Graph that records dependency, co-usage, and conflict edges among skills, co-evolves related skills when one changes, clusters trace-specific proposals into general capabilities, and replays historical cases before accepting a new skill. Evaluated on bug-revealing test generation and false-positive bug report filtering with two coding agents, GSE reports consistently higher precision, recall, and F1 than local-update baselines, and a deployment on an industrial agent improves F1 from 0.43 to 0.71. If the claim holds, coding agents can keep improving across projects without expensive retraining.
What carries the argument. The load-bearing mechanism is the Skill Relation Graph (SRG), a directed graph whose nodes are skills and whose edges carry one of three semantic relation types: dependency (one skill relies on assumptions or outputs of another), co-usage (two skills work best together), and conflict (two skills give incompatible guidance). The SRG lets the update agent reason about which skills a proposed change affects, generate auxiliary proposals to keep related skills consistent, and evolve the relation edges themselves as first-class objects. Around the graph sit two generalization components: cluster-based consolidation, which groups proposals by affected skill set or by semantic similarity of new-skill rationales and merges each cluster into a higher-level capability, and replay-driven validation, which accepts a candidate skill bank only if replaying relevant historical cases keeps or improves performance. Together they convert trace-specific lessons into reusable skills while filtering out overfitted or regression-prone updates.
Core claim. The paper's central claim is that the skill bank itself is the unit of evolution, not individual skills. GSE formulates skill evolution as a global optimization over an interconnected bank: a Skill Relation Graph (SRG) models dependency, co-usage, and conflict relations, and every proposed update is checked against neighboring skills and propagated through auxiliary proposals when it changes their assumptions. Locally successful proposals are then clustered by affected skills or semantic similarity, merged into higher-level abstractions, and admitted to the bank only if replay on relevant historical cases preserves or improves performance. On 108 Java bugs from Multi-SWE-Bench and 500 Go bug reports from an industrial dataset, GSE reports F1 improvements over the strongest baselines of 31.0% to 93.8% on test generation and 17.9% to 66.7% on false-positive filtering, across both OpenHands and mini-SWE-agent.
The premise it rests on. GSE depends on the language model correctly judging which skills depend on, complement, or contradict one another, and on replaying a small retrieved set of historical cases being a reliable signal that a new skill will generalize to unseen projects.
If the paper is right (Pith's inferences, not the authors')
Skills evolved on one set of projects transfer to held-out projects, since the one-project-held-out protocol keeps evaluation projects out of the evolution process.
Skill banks can grow without accumulating contradictory rules: conflicting updates to stateful versus stateless test construction are resolved into a dependency between a general diagnostic skill and the existing construction skills.
The best-performing skill bank also costs fewer execution-time tokens than human-written skills or local-update evolution methods, so globalized evolution is not bought with runtime overhead.
Stronger, already-tuned industrial agents still gain from globalized skill evolution, improving F1 from 0.43 to 0.71 on false-positive filtering.
Both the graph and the generalization mechanism are individually necessary: ablations removing either component lower F1 in both tasks.
A testable extension is to measure SRG edge accuracy directly: annotate dependency, co-usage, and conflict relations by hand for a sample of the skill bank and compare, because noisy edges that happen not to be exercised by replay cases would be invisible to the current validation.
The clustering key for existing-skill proposals is the affected skill set; generalizing beyond that, for example clustering on shared rationale across different affected skills, might find cross-skill capabilities the current design misses.
Replay sets are retrieved historical cases rather than a sample from the full task distribution, so the generalization guarantee is only as strong as retrieval; replaying a broader or adversarially selected set of cases would test whether the guarantee holds more widely.
If the approach scales, skill evolution could be applied to other agent behaviors beyond software engineering, such as tool-use or data-analysis workflows, wherever a bank of procedural skills and historical traces exists.
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
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.
The paper
Abstract
Automated skill evolution enables Large Language Model (LLM) agents to continuously improve without expensive retraining. However, existing approaches typically treat skill evolution as a sequence of local updates, overlooking relationships among skills and often producing overfitted skill updates that fail to generalize across tasks. We propose GSE, a globalized skill evolution framework that jointly optimizes skill compatibility and skill generalization. To preserve consistency across the skill bank, GSE maintains a Skill Relation Graph (SRG) that explicitly models and co-evolves inter-skill relationships. To improve generalization, GSE performs cluster-based skill consolidation to abstract reusable capabilities from local updates and employs replay-driven verification to prevent overfitting and behavioral regressions. We evaluate GSE on two representative software engineering tasks: bug-revealing test generation and false-positive bug report filtering. Across two state-of-the-art coding agents, OpenHands and mini-SWE-agent, GSE consistently achieves the best precision, recall, and F1-score. Compared with existing evolution techniques, GSE improves precision and recall by 6.1% 34.1% and 31.8% 180.0% for test generation, and by 15.4% 96.4% and 13.1% 19.8% for false-positive filtering. Deployment on an internal industrial agent further yields a 61.4% improvement in F1-score, demonstrating the effectiveness and generalizability of GSE for evolving effective skills.
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 · 2026-09-23
The authors' main claim, as the classifier picked it out 0.97
“Deployment on an internal industrial agent further yields a 61.4% improvement in F1-score, demonstrating the effectiveness and generalizability of GSE for evolving effective skills.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.07 · human participants 0.03 · controlled comparison 0.86 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03
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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Automated skill evolution enables Large Language Model (LLM) agents to continuously improve without expensive retraining. | background 1.00 | 0.07 | 0.05 | asserted 1.00 |
| However, existing approaches typically treat skill evolution as a sequence of local updates, overlooking relationships among skills and often producing overfitted skill updates that fail to generalize across tasks. | gap 0.96 | 0.06 | 0.03 | asserted 0.98 |
| We propose GSE, a globalized skill evolution framework that jointly optimizes skill compatibility and skill generalization. | contribution 0.99 | 0.89 | 0.12 | asserted 0.99 |
| To preserve consistency across the skill bank, GSE maintains a Skill Relation Graph (SRG) that explicitly models and co-evolves inter-skill relationships. | method 0.98 | 0.85 | 0.03 | asserted 1.00 |
| To improve generalization, GSE performs cluster-based skill consolidation to abstract reusable capabilities from local updates and employs replay-driven verification to prevent overfitting and behavioral regressions. | method 0.98 | 0.88 | 0.06 | asserted 1.00 |
| We evaluate GSE on two representative software engineering tasks: bug-revealing test generation and false-positive bug report filtering. | method 0.76 | 0.95 | 0.03 | asserted 0.97 |
| Across two state-of-the-art coding agents, OpenHands and mini-SWE-agent, GSE consistently achieves the best precision, recall, and F1-score. | result 1.00 | 0.94 | 0.25 | asserted 1.00 |
| Compared with existing evolution techniques, GSE improves precision and recall by 6.1%~34.1% and 31.8%~180.0% for test generation, and by 15.4%~96.4% and 13.1%~19.8% for false-positive filtering. | result 1.00 | 0.95 | 0.19 | asserted 1.00 |
| Deployment on an internal industrial agent further yields a 61.4% improvement in F1-score, demonstrating the effectiveness and generalizability of GSE for evolving effective skills. | result 0.94 | 0.95 | 0.60 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99
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, 2026-09-23: 1 request, 7,064 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 (3)
References (37)
- [1] Anthropic skills, 2026.https://github.com/anthropics/skills on Pith
- [2] deepseek, 2026.https://www.deepseek.com on Pith
- [3] langchain, 2026.https://github.com/langchain-ai/langchain on Pith
- [4] Swe-bench leaderboard, 2026.https://www.swebench.com/index.html on Pith
- [5] Hongcheol Cho, Ryangkyung Kang, and Youngeun Kim. Skillret: A large-scale benchmark for skill retrieval in llm agents, 2026 on Pith
- [6] Elizabeth Dinella, Gabriel Ryan, Todd Mytkowicz, and Shuvendu K Lahiri. Toga: A neural method for test oracle generation. InProceedings of the 44th International Conference on Software Engineering, pages 2130–2141, 2022 on Pith
- [7] Xueying Du, Jiayi Feng, Yi Zou, Wei Xu, Jie Ma, Wei Zhang, Sisi Liu, Xin Peng, and Yiling Lou. Reducing false positives in static bug detection with llms: An empirical study in industry, 2026 on Pith
- [8] 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 on PithPith's reviewarXiv
- [9] Jiazhan Feng, Shijue Huang, Xingwei Qu, Ge Zhang, Yujia Qin, Baoquan Zhong, Chengquan Jiang, Jinxin Chi, and Wanjun Zhong. Retool: Reinforcement learning for strategic tool use in llms, 2025 on Pith
- [10] 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?arXiv preprint arXiv:2603.15401, 2026 on PithPith's reviewarXiv
- [11] Yoonho Lee, Roshen Nair, Qizheng Zhang, Kangwook Lee, Omar Khattab, and Chelsea Finn. Meta-harness: End-to-end optimization of model harnesses, 2026 on Pith
- [12] 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, 2026 on Pith
- [13] Yanzhou Li, Yiran Zhang, Xiaoyu Zhang, Xiaoxia Liu, and Yang Liu. Codeskill: Learning self-evolving skills for coding agents, 2026 on Pith
- [14] 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, 2026 on Pith
- [15] Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T Le, Samira Daruki, Xiangru Tang, et al. Reasoningbank: Scaling agent self-evolving with reason- ing memory.arXiv preprint arXiv:2509.25140, 2025 on PithPith's reviewarXiv
- [16] Zehan Qi, Xiao Liu, Iat Long Iong, Hanyu Lai, Xueqiao Sun, Jiadai Sun, Xinyue Yang, Yu Yang, Shuntian Yao, Wei Xu, et al. Webrl: Training llm web agents via self-evolving online curriculum reinforcement learning, 2025 on Pith
- [17] Timo Schick, Jane Dwivedi-Yu, Roberto Dess`ı, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.Advances in neural information processing systems, 36:68539– 68551, 2023 on Pith
- [18] Yaorui Shi, Yuxin Chen, Zhengxi Lu, Yuchun Miao, Shugui Liu, Qi Gu, Xunliang Cai, Xiang Wang, and An Zhang. Skill1: Unified evolution of skill-augmented agents via reinforcement learning.arXiv preprint arXiv:2605.06130, 2026. 16 on PithPith's reviewarXiv
- [19] Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Re- flexion: Language agents with verbal reinforcement learning.Advances in neural information processing systems, 36:8634–8652, 2023 on Pith
- [20] Jiashuo Tian, Dong Wang, Chen Yang, Haichi Wang, Zan Wang, and Junjie Chen. Character- izing and mitigating false-positive bug reports in the linux kernel, 2026 on Pith
show 17 more
- [21] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023 on PithPith's reviewarXiv
- [22] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Openhands: An open platform for ai software developers as generalist agents. InInternational Conference on Learning Representations, volume 20... on Pith
- [23] Yinjie Wang, Xuyang Chen, Xiaolong Jin, Mengdi Wang, and Ling Yang. Openclaw-rl: Train any agent simply by talking, 2026 on Pith
- [24] Chunqiu Steven Xia, Matteo Paltenghi, Jia Le Tian, Michael Pradel, and Lingming Zhang. Fuzz4all: Universal fuzzing with large language models. InProceedings of the IEEE/ACM 46th International Conference on Software Engineering, pages 1–13, 2024 on Pith
- [25] Chunqiu Steven Xia, Zhe Wang, Yan Yang, Yuxiang Wei, and Lingming Zhang. Live- swe-agent: Can software engineering agents self-evolve on the fly?arXiv preprint arXiv:2511.13646, 2025 on PitharXiv
- [26] Chunqiu Steven Xia, Zhe Wang, Yan Yang, Yuxiang Wei, and Lingming Zhang. Live-swe- agent: Can software engineering agents self-evolve on the fly?, 2025 on Pith
- [27] 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 on PithPith's reviewarXiv
- [28] Renjun Xu and Yang Yan. Agent skills for large language models: Architecture, acquisition, security, and the path forward.arXiv preprint arXiv:2602.12430, 2026 on PithPith's reviewarXiv
- [29] Chen Yang and Junjie Chen. Uncovering business logic bugs via semantics-driven unit test generation, 2026 on Pith
- [30] Chen Yang, Ziqi Wang, Yanjie Jiang, Lin Yang, Yuteng Zheng, Jianyi Zhou, and Junjie Chen. Reflective unit test generation for precise type error detection with large language models. arXiv preprint arXiv:2507.02318, 2025 on PitharXiv
- [31] John Yang, Carlos Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering.Advances in Neural Information Processing Systems, 37:50528–50652, 2024 on Pith
- [32] Xin Yin, Chao Ni, Xiaodan Xu, and Xiaohu Yang. What you see is what you get: Attention- based self-guided automatic unit test generation. In2025 IEEE/ACM 47th International Con- ference on Software Engineering (ICSE), pages 1039–1051. IEEE, 2025 on Pith
- [33] Daoguang Zan, Zhirong Huang, Wei Liu, Hanwu Chen, Shulin Xin, Linhao Zhang, Qi Liu, Li Aoyan, Lu Chen, Xiaojian Zhong, et al. Multi-swe-bench: A multilingual benchmark for issue resolving, 2026 on Pith
- [34] Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow gener- ation. InInternational Conference on Learning Representations, volume 2025, pages 34040– 34077, 2025 on Pith
- [35] Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel: Llm agents are experiential learners. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632–19642, 2024. 17 on Pith
- [36] YanZhao Zheng, ZhenTao Zhang, Chao Ma, YuanQiang Yu, JiHuai Zhu, Yong Wu, Tianze Xu, Baohua Dong, Hangcheng Zhu, Ruohui Huang, and Gang Yu. Skillrouter: Skill routing for llm agents at scale, 2026 on Pith
- [37] Huichi Zhou, Yihang Chen, Siyuan Guo, Xue Yan, Kin Hei Lee, Zihan Wang, Ka Yiu Lee, Guchun Zhang, Kun Shao, Linyi Yang, and Jun Wang. Memento: Fine-tuning llm agents without fine-tuning llms, 2025. 18 on Pith
Cited by 2 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
How the literature uses it
1 arXiv paper cite this in 3 sentences
What it cites · 20 sentences without a confident label
Uses the method or tool of 1 paper
-
For the bug-revealing test generation task, we also follow existing work [ref] and check whether any bug-triggering tests generated by match the benchmark reference tests.
2507.02318 · Threats to Validity · 0.86
Compares with 2 papers
-
For each, we compare five settings: the vanilla agent, the agent augmented with human-written skills, Live-SWE-agent [ref] , Trace2Skill [ref] , and .
2511.13646 · Introduction · 0.95
-
For each, we compare five settings: the vanilla agent, the agent augmented with human-written skills, Live-SWE-agent [ref] , Trace2Skill [ref] , and .
Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills · Introduction · 0.95
Cites as background 10 papers
-
Recent efforts further formalize this paradigm into comprehensive skill systems, where skills are treated as explicit artifacts that encapsulate reusable reasoning strategies, decision procedures, and task-specific knowledge for on-demand use [ref] .
Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward · Related Work · 0.87
-
Recent agentic reinforcement learning methods optimize agent policies for long-horizon reasoning and interaction tasks, such as ReTool [ref] , WebRL [ref] , and OpenClaw-RL [ref] , which train agents to better utilize tools and interact with environments through reinforcement learning.
ReTool: Reinforcement Learning for Strategic Tool Use in LLMs · Related Work · 0.99
-
Recent agentic reinforcement learning methods optimize agent policies for long-horizon reasoning and interaction tasks, such as ReTool [ref] , WebRL [ref] , and OpenClaw-RL [ref] , which train agents to better utilize tools and interact with environments through reinforcement learning.
OpenClaw-RL: Train Any Agent Simply by Talking · Related Work · 0.99
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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.



No comments yet.
Sign in to comment