pith. sign in

arxiv: 2606.30015 · v1 · pith:GKQRTUJZnew · submitted 2026-06-29 · 💻 cs.CL

Parametric Skills

Xuan Zhao , Haonan He , Qingyu Yang , Minglei Li , Jingqi Ye , Zelin Tan , Bo Wan , Peng Ye This is my paper

Pith reviewed 2026-06-30 06:18 UTC · model grok-4.3

classification 💻 cs.CL
keywords ParametricSkillshypernetworkLoRA adaptersskill exploitationsoftware engineering tasksin-context learningcontinual learningagent capabilities
0
0 comments X

The pith

A hypernetwork converts textual skills into LoRA adapters at test time, enabling context-free exploitation that outperforms in-context learning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to fix the problem that large language models struggle to locate and follow long or complex skill instructions during agentic tasks. It builds a skill library, generates single-turn and multi-turn exploitation trajectories, and trains a hypernetwork that ingests any textual skill and emits a corresponding LoRA adapter. The resulting adapters are inserted into the model so the skill can be used without its text remaining in context. Experiments on six software-engineering subtasks show average gains of 6.44 points over in-context learning plus higher BERT and F1 scores. The same mechanism is presented as a route to accumulative, test-time continual learning.

Core claim

ParametricSkills trains a hypernetwork on synthesized trajectories so that any free-form textual skill can be turned into a LoRA adapter; once inserted, the adapter lets the model exploit the skill without the original text in context, producing stronger results than in-context learning on complex software-engineering subtasks while also supporting additive skill accumulation over time.

What carries the argument

The hypernetwork that maps a textual skill description to a set of LoRA adapter weights for insertion into the base model.

If this is right

  • Skills become usable without occupying context length or requiring precise instruction following.
  • Average performance on the six SWE subtasks rises 6.44 points relative to in-context learning.
  • BERT Score and F1 both increase, indicating both semantic and surface-level improvements.
  • Skills can be added incrementally at test time without retraining the base model.

Where Pith is reading between the lines

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

  • Because the adapters are additive, the same hypernetwork could support skill libraries that grow across sessions or users without interference.
  • The separation of skill content from the base model weights may allow the same learned adapters to be reused or composed on different base models.
  • If the hypernetwork generalizes, new skills could be created from a few demonstrations and immediately turned into adapters without manual rewriting.

Load-bearing premise

The single-turn and multi-turn trajectories synthesized with OpenCode around the collected skills are representative of the exploitation patterns that will appear at test time.

What would settle it

A clear drop in the reported gains when the same hypernetwork is tested on software-engineering problems whose structure or distribution differs markedly from the OpenCode-synthesized training set.

Figures

Figures reproduced from arXiv: 2606.30015 by Bo Wan, Haonan He, Jingqi Ye, Minglei Li, Peng Ye, Qingyu Yang, Xuan Zhao, Zelin Tan.

Figure 1
Figure 1. Figure 1: Overall Pipeline of ParametricSkills. (I) The parametric skill generation pipeline of ParametricSkills; (II)&(III) The training pipeline of ParametricSkills, from pretraining to multi-turn skill-exploitation fine-tuning; (IV) The self-evolving pipeline and continual learning pipeline of ParametricSkills. 2026; Wang et al., 2026a), the lack of sufficient capability in understanding and following skills may … view at source ↗
Figure 2
Figure 2. Figure 2: Training Data Construction Pipeline of ParametricSkills. files, etc. (2) experience-driven skills mined from successful agent execution trajectories. These trajectories provide complete task-closure evidence, including problem localization, code inspection, and so on. To convert raw agent trajectories into a unified skill representa￾tion, we design a three-stage short-call pipeline. First, the Summarize st… view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of Categories of Collected Skills. After skill library curation, we construct single￾turn skill-exploitation samples for each skill, where the model must correctly use the skills provided with related tasks. For each skill, we construct two complementary types of single-turn samples: (1) short QA-style samples designed to capture specific, narrowly scoped skill-use sce￾narios. These samples ar… view at source ↗
Figure 4
Figure 4. Figure 4: presents an example of our structured skill definition template. The skill is organized into several functional components, including the skill purpose, usage conditions, required inputs, implementation recipe, verification checks, failure modes and anti-patterns. This structure is designed to capture reusable procedural knowledge rather than a task-specific solution. Skill Card: Implement API Resource Lis… view at source ↗
Figure 5
Figure 5. Figure 5: Pretraining Loss and PPL Curves 0 5000 10000 15000 20000 25000 30000 35000 Step 0.4 0.5 0.6 0.7 0.8 0.9 Loss Single-Turn Skill-Expolitation Trajectory Fine-Tune Loss train/tmp_loss raw train/tmp_loss smoothed eval/loss best @ 35000 epoch 1 end: 17513 epoch 2 end: 35026 0 5000 10000 15000 20000 25000 30000 35000 Step 1.6 1.8 2.0 2.2 2.4 2.6 PPL Single-Turn Skill-Expolitation Trajectory Fine-Tune PPL train/t… view at source ↗
Figure 6
Figure 6. Figure 6: Single-turn Skill Exploitation Fine-tuning Loss and PPL Curves [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Multi-turn Skill Exploitation Fine-tuning Loss and PPL Curves [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: illustrates the construction of a skill-augmented HumanEval instance. The upper block represents the SKILL.md context, which contains reusable procedural knowledge distilled from a verified successful trajectory, while the lower block keeps the original HumanEval problem prompt unchanged. During evaluation, the model receives the concatenation of the structured skill document and the HumanEval prompt, and … view at source ↗
Figure 9
Figure 9. Figure 9: presents the prompt template and 0–100 scoring rubric used by the LLM judge, where DeepSeek-V4-Flash serves as the judge model. The judge compares each predicted answer against the ground-truth answer as the semantic reference, emphasizing semantic correctness rather than surface-form similarity. The scoring rubric further penalizes unsupported details, fabricated constraints, missing key requirements, and… view at source ↗
read the original abstract

Since intelligence fundamentally relies on efficient skill acquisition (Chollet, 2019), the ability to leverage skills is critical. For LLMs, skills, manually authored or extracted from task trajectories, are textual recipes encoding mature problem-solving experience and are critical to agentic capabilities. Despite widespread deployment, their utility is limited by the model's ability to comprehend and follow skill instructions, especially under complex and long-context scenarios, where key instructions are difficult to locate and adhere to. To address this limitation, we propose ParametricSkills, a framework that can convert free-form textual skills into parameters at test time, enabling context-free skill exploitation. Specifically, we first construct a large-scale, high-quality skill library, and synthesize single-turn and multi-turn skill exploitation trajectories built around these skills with OpenCode. Using these data, we then train a hypernetwork that parameterizes both the skill content and the test-time exploitation methodology by receiving textual skills and converting them into LoRA adapters. Experimental results on six complex software engineering (SWE) subtasks demonstrate that, the proposed ParametricSkills averagely outperforms in-context learning by 6.44 points as judged by DeepSeek-V4-Flash, while also achieving significantly higher BERT Score and F1 score, confirming its effectiveness. Beyond performance, we further find that parametric skills, being inherently accumulative, offer a preliminary yet promising avenue toward test-time continual learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes ParametricSkills, a framework that trains a hypernetwork to convert free-form textual skills into LoRA adapters at test time. Skills are drawn from a constructed library; single- and multi-turn exploitation trajectories are synthesized with OpenCode; the hypernetwork is trained on these trajectories so that skill content and exploitation methodology are both parameterized. On six software-engineering subtasks the method is reported to outperform in-context learning by an average of 6.44 points according to DeepSeek-V4-Flash judgments, while also showing higher BERT Score and F1; the authors further note that the parametric representation is inherently accumulative and therefore offers a route to test-time continual learning.

Significance. If the performance advantage can be shown to reflect genuine task improvement rather than evaluator artifact, the work would supply a concrete mechanism for making skills context-free and accumulative, addressing a recognized bottleneck in long-context agentic behavior. The hypernetwork approach to skill parameterization is technically distinct from standard in-context or retrieval-based methods and, if reproducible, would constitute a measurable advance in the direction of continual, parameter-level skill acquisition.

major comments (2)
  1. [Abstract] Abstract: the headline claim of a 6.44-point average improvement rests exclusively on judgments produced by DeepSeek-V4-Flash. No calibration against human labels, inter-annotator agreement, or correlation with execution-based correctness is reported. Because the judge is the sole source of the numeric superiority, any systematic preference for the LoRA-injected output style or for the syntactic patterns of OpenCode-synthesized trajectories would render the central performance claim uninterpretable.
  2. [Abstract] Abstract and experimental description: the training trajectories are synthesized separately via OpenCode, yet no analysis is provided of how well these trajectories cover the distribution of real test-time exploitation scenarios. The weakest assumption—that the synthetic single- and multi-turn data suffice for generalization—is therefore load-bearing for the reported gains, but remains untested.
minor comments (2)
  1. [Abstract] The abstract states that ParametricSkills also achieves “significantly higher BERT Score and F1 score,” but does not indicate whether these metrics correlate with the LLM-judge scores or with any execution-based ground truth.
  2. [Abstract] No information is given on the number of runs, variance, or statistical tests supporting the 6.44-point figure.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of a 6.44-point average improvement rests exclusively on judgments produced by DeepSeek-V4-Flash. No calibration against human labels, inter-annotator agreement, or correlation with execution-based correctness is reported. Because the judge is the sole source of the numeric superiority, any systematic preference for the LoRA-injected output style or for the syntactic patterns of OpenCode-synthesized trajectories would render the central performance claim uninterpretable.

    Authors: We agree that the 6.44-point figure depends on the LLM judge and that explicit calibration to human labels or execution-based metrics is not reported. At the same time, the manuscript already presents significantly higher BERT Score and F1 as orthogonal, automatic metrics that do not rely on the judge. These results indicate improved semantic similarity and precision independent of any stylistic bias. We will revise the abstract and results section to foreground the automatic metrics, add a limitations paragraph discussing possible judge biases, and qualify the headline claim accordingly. revision: partial

  2. Referee: [Abstract] Abstract and experimental description: the training trajectories are synthesized separately via OpenCode, yet no analysis is provided of how well these trajectories cover the distribution of real test-time exploitation scenarios. The weakest assumption—that the synthetic single- and multi-turn data suffice for generalization—is therefore load-bearing for the reported gains, but remains untested.

    Authors: The single- and multi-turn trajectories were generated with OpenCode specifically to produce exploitation sequences grounded in the skill library for software-engineering contexts. While we did not conduct an explicit distributional overlap analysis between the synthetic data and real test-time scenarios, the consistent gains across six diverse subtasks provide indirect support for generalization. We will expand the experimental section with additional detail on the synthesis procedure and add a limitations discussion that explicitly notes the untested coverage assumption. revision: partial

Circularity Check

0 steps flagged

No circularity; empirical gains rest on held-out evaluation after separate synthesis

full rationale

The paper describes constructing a skill library, synthesizing trajectories via OpenCode, training a hypernetwork to produce LoRA adapters from textual skills, and then reporting average gains of 6.44 points (plus BERT/F1) versus in-context learning on six SWE subtasks. No equations, fitted-parameter renamings, or self-citation chains appear that would make the reported performance reduce to the training inputs by construction. The evaluation is on held-out tasks with metrics external to the hypernetwork training loop, rendering the central claim self-contained rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, mathematical axioms, or invented entities with independent evidence; the hypernetwork is introduced as the core mechanism without further decomposition.

pith-pipeline@v0.9.1-grok · 5787 in / 1214 out tokens · 37193 ms · 2026-06-30T06:18:56.223832+00:00 · methodology

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

29 extracted references · 27 canonical work pages · 20 internal anchors

  1. [1]

    EvoSkill: Automated Skill Discovery for Multi-Agent Systems

    Salaheddin Alzubi, Noah Provenzano, Jaydon Bingham, Weiyuan Chen, and Tu Vu. Evoskill: Automated skill discovery for multi-agent systems.arXiv preprint arXiv:2603.02766,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

    Accessed: 2026-06-26. Rujikorn Charakorn, Edoardo Cetin, Yujin Tang, and Robert Tjarko Lange. Text-to-lora: Instant transformer adaption.arXiv preprint arXiv:2506.06105,

    work page arXiv 2026

  3. [3]

    arXiv preprint arXiv:2602.15902 , year =

    Rujikorn Charakorn, Edoardo Cetin, Shinnosuke Uesaka, and Robert Tjarko Lange. Doc- to-lora: Learning to instantly internalize contexts.arXiv preprint arXiv:2602.15902,

    work page arXiv

  4. [4]

    On the Measure of Intelligence

    François Chollet. On the measure of intelligence.arXiv preprint arXiv:1911.01547,

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  5. [5]

    Accessed: 2026-06-23

    URL https://github.com/sst/opencode. Accessed: 2026-06-23. DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence,

    2026

  6. [6]

    The Llama 3 Herd of Models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    HyperNetworks

    David Ha, Andrew Dai, and Quoc V Le. Hypernetworks.arXiv preprint arXiv:1609.09106,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Winning gold at imo 2025 with a model-agnostic verification- and-refinement pipeline.arXiv preprint arXiv:2507.15855, 2025

    Yichen Huang and Lin F Yang. Winning gold at imo 2025 with a model-agnostic verification- and-refinement pipeline.arXiv preprint arXiv:2507.15855,

    work page arXiv 2025

  9. [9]

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling

    Yafu Li, Runzhe Zhan, Haoran Zhang, Shunkai Zhang, Yizhuo Li, Zhilin Wang, Jiacheng Chen, Futing Wang, Xuyang Hu, Yuchen Fan, et al. Achieving gold-medal-level olympiad reasoning via simple and unified scaling.arXiv preprint arXiv:2605.13301,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Se-agent: Self-evolution trajectory optimization in multi-step reasoning with llm-based agents.arXiv preprint arXiv:2508.02085, 2025

    Jiaye Lin, Yifu Guo, Yuzhen Han, Sen Hu, Ziyi Ni, Licheng Wang, Mingguang Chen, Hongzhang Liu, Ronghao Chen, Yangfan He, et al. Se-agent: Self-evolution trajectory op- timization in multi-step reasoning with llm-based agents.arXiv preprint arXiv:2508.02085,

    work page arXiv

  12. [12]

    A comprehensive survey on long context language modeling.arXiv preprint arXiv:2503.17407,

    Jiaheng Liu, Dawei Zhu, Zhiqi Bai, Yancheng He, Huanxuan Liao, Haoran Que, Zekun Wang, Chenchen Zhang, Ge Zhang, Jiebin Zhang, et al. A comprehensive survey on long context language modeling.arXiv preprint arXiv:2503.17407,

    work page arXiv

  13. [13]

    SHINE: A Scalable In-Context Hypernetwork for Mapping Context to LoRA in a Single Pass

    9 Preprint. Under review. Yewei Liu, Xiyuan Wang, Yansheng Mao, Yoav Gelbery, Haggai Maron, and Muhan Zhang. Shine: A scalable in-context hypernetwork for mapping context to lora in a single pass. arXiv preprint arXiv:2602.06358,

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    RoBERTa: A Robustly Optimized BERT Pretraining Approach

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692,

    work page internal anchor Pith review Pith/arXiv arXiv 1907

  15. [15]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization

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

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Humanity's Last Exam

    Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning

    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,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    ALFWorld: Aligning Text and Embodied Environments for Interactive Learning

    Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté, Yonatan Bisk, Adam Trischler, and Matthew Hausknecht. Alfworld: Aligning text and embodied environments for interactive learning.arXiv preprint arXiv:2010.03768,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  20. [20]

    Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents

    Hao Wang, Guozhi Wang, Han Xiao, Yufeng Zhou, Yue Pan, Jichao Wang, Ke Xu, Yafei Wen, Xiaohu Ruan, Xiaoxin Chen, et al. Skill-sd: Skill-conditioned self-distillation for multi-turn llm agents.arXiv preprint arXiv:2604.10674, 2026a. Miles Wang, Robi Lin, Kat Hu, Joy Jiao, Neil Chowdhury, Ethan Chang, and Tejal Pat- wardhan. Frontierscience: Evaluating ai’s...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

    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,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    Lora-gen: Specializing large language model via online lora generation.arXiv preprint arXiv:2506.11638,

    Yicheng Xiao, Lin Song, Rui Yang, Cheng Cheng, Yixiao Ge, Xiu Li, and Ying Shan. Lora-gen: Specializing large language model via online lora generation.arXiv preprint arXiv:2506.11638,

    work page arXiv

  23. [23]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    ARIS: Autonomous Research via Adversarial Multi-Agent Collaboration

    Ruofeng Yang, Yongcan Li, and Shuai Li. Aris: Autonomous research via adversarial multi-agent collaboration.arXiv preprint arXiv:2605.03042, 2026a. 10 Preprint. Under review. 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.ar...

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    Hipho: How far are (m) llms from humans in the latest high school physics olympiad benchmark?arXiv preprint arXiv:2509.07894,

    Fangchen Yu, Haiyuan Wan, Qianjia Cheng, Yuchen Zhang, Jiacheng Chen, Fujun Han, Yulun Wu, Junchi Yao, Ruilizhen Hu, Ning Ding, et al. Hipho: How far are (m) llms from humans in the latest high school physics olympiad benchmark?arXiv preprint arXiv:2509.07894,

    work page arXiv

  26. [26]

    GLM-5: from Vibe Coding to Agentic Engineering

    Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification

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

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications

    Yingli Zhou, Wang Shu, Yaodong Su, Wenchuan Du, Yixiang Fang, and Xuemin Lin. A comprehensive survey on agent skills: Taxonomy, techniques, and applications.arXiv preprint arXiv:2605.07358,

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    POST" and id=None. cls.create(attach_host_name=False, method=

    and the search- augmented question-answering benchmark Search-QA (Jin et al., 2025).ParametricSkills shares the skill-to-LoRA backbone but differs in three aspects: (i) we target real-world agentic tasks such as production-level software engineering rather than a simple demonstration on ALFWorld and Search-QA, (ii) we construct a comprehensive skills libr...

    2025