CoChat
Sign Up
PreprintCited by 9Web Data Mining and AnalysisInformation Retrieval and Search BehaviorAI-based Problem Solving and Planning

SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills

Open access
Zheng, Boyuan, Fatemi, Michael Y., Jin, Xiaolong, Wang, Zora Zhiruo, Gandhi, Apurva, Song, Yueqi, Gu, Yu, Srinivasa, Jayanth +3 more·2025·arXiv
DOI 10.48550/arxiv.2504.07079arXiv 2504.07079
Open this paper in CoChat — free accountOpen PDF

Abstract

To survive and thrive in complex environments, humans have evolved sophisticated self-improvement mechanisms through environment exploration, hierarchical abstraction of experiences into reuseable skills, and collaborative construction of an ever-growing skill repertoire. Despite recent advancements, autonomous web agents still lack crucial self-improvement capabilities, struggling with procedural knowledge abstraction, refining skills, and skill composition. In this work, we introduce SkillWeaver, a skill-centric framework enabling agents to self-improve by autonomously synthesizing reusable skills as APIs. Given a new website, the agent autonomously discovers skills, executes them for practice, and distills practice experiences into robust APIs. Iterative exploration continually expands a library of lightweight, plug-and-play APIs, significantly enhancing the agent's capabilities. Experiments on WebArena and real-world websites demonstrate the efficacy of SkillWeaver, achieving relative success rate improvements of 31.8% and 39.8%, respectively. Additionally, APIs synthesized by strong agents substantially enhance weaker agents through transferable skills, yielding improvements of up to 54.3% on WebArena. These results demonstrate the effectiveness of honing diverse website interactions into APIs, which can be seamlessly shared among various web agents.

AI snapshot

Written from the paper’s full text.

TL;DR

SkillWeaver is a skill-centric framework that lets web agents self-improve by autonomously proposing, practicing, synthesizing, and debugging reusable Playwright Python APIs for websites. Across WebArena and live websites, adding these synthesized APIs improves agent success rates and transfers to weaker agents.

Key findings

  • On WebArena, SkillWeaver APIs improved GPT-4o average success from 22.6% to 29.8% (↑32% relative) and GPT-4o-mini from 9.2% to 14.1% (↑45% relative); the abstract/conclusion summarize WebArena relative gains as 31.8% and weaker-agent improvements up to 54.3%.
  • On a live Online-Mind2Web subset (4 websites, 57 tasks), average success rose from 40.2% to 56.2% (↑40% relative); the abstract/conclusion report a 39.8% relative improvement on real-world websites.
  • Synthesized APIs performed comparably to human-crafted APIs on low/medium API-support sites such as Reddit and Shopping, but not as well on high API-support sites such as GitLab and Maps.
  • Compositional APIs emerged, calling multiple simpler APIs, indicating increasing abstraction and efficiency in synthesized skills.

Methods

  • Three-stage pipeline: Skill Proposal (LLM proposes procedural, navigational, and information-seeking tasks), Skill Synthesis (practice tasks, use an LLM reward model, generate reusable Playwright Python APIs), and Skill Honing (unit-test and debug APIs).
  • Evaluation on WebArena (812 tasks across Gitlab, Map, Shopping, CMS, Reddit) and live websites from Online-Mind2Web (4 websites, 57 tasks); exploration used 160 iterations per website with GPT-4o.
  • Baseline agent uses Code-Act with GPT-4o at temperature 0.3 and max 10 steps; weaker agent swaps in GPT-4o-mini; an API selection module filters relevant APIs. WebArena uses automated metrics; live website trajectories were manually evaluated.

Limitations

  • The paper states LLMs are not robust enough at API calling: failures include not identifying the appropriate API and generating wrong parameters, with weaker LLMs such as GPT-4o-mini more affected.
  • The paper reports synthesized APIs underperform human-crafted APIs on high API-support websites (GitLab, Maps), and on the live Car website final success rate remained unchanged even though APIs reached the required final state in 4 of 9 tasks.
  • The paper notes its verification criterion—calling an API without producing an exception—can mark malfunctioning APIs as verified if they silence exceptions; it also flags safety concerns for live websites and manually validated all actions before execution.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

DataCite · 10.48550/arxiv.2504.07079

Cites

0 citation errors · 1 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

4 of 13 cited sources checked · 2 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Natural language-based routines pose challenges for formal verification and precise composition into new workflows.

    “However, its natural language-based routines pose challenges for formal verification and precise composition into new workflows.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Wang et al., 2024e

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Sarch et al., 2024 source

      Section 3.2 (pp. 4–5) describes natural-language thoughts and task plans, while Section 3.3 (p. 5) describes verification through execution with a human observer. Neither passage says these routines pose challenges for formal verification or precise composition into new workflows; searches for those concepts found no supporting passage in our copy. (LLM, conf=0.72)

      Experimental — accuracy not yet measured
    • Rests on an assumption: Natural language-based routines pose challenges for formal verification and precise composition into new workflows.
  • Findingown data

    After a certain number of exploration iterations, the pipeline begins to generate compositional APIs that call multiple simpler APIs.

    “After a certain number of exploration iterations, we observe that the pipeline begins to generate compositional APIs that call multiple simpler APIs.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.
    • Rests on an assumption: LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.
  • Findingown data

    With weaker agents, relative improvements range from 40% to 130%.

    “Further experiments with weaker agents revealed even more dramatic improvements, ranging from 40% to 130%.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: WebArena is a good platform for demonstrating the efficacy of the agent.
  • Findingown data

    Synthesized APIs perform comparably to human-crafted APIs on websites with low API support (Reddit) and medium API support (Shopping).

    “the performance of the synthesized APIs is comparable with human-crafted APIs on websites with low API support, such as Reddit, and those with medium API support, like Shopping.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Human-crafted APIs extracted from official documentation are a fair comparison baseline for synthesized APIs.
  • Findingown data

    On WebArena, GPT-4o-mini with synthesized APIs shows relative success rate improvements ranging from 40% to 133% across five websites.

    “upon integration with our synthesized APIs, the agent based on GPT-4o-mini demonstrates remarkable relative success rate improvements ranging from 40% to 133% across five websites.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: WebArena is a good platform for demonstrating the efficacy of the agent.
  • May combine two claims. 15
  • Classified as probably not a substantial claim. 21
  • The quote only partly states this claim. 1

  • May combine two claims.

    Premisecitation

    Agent Workflow Memory and ICAL focus on online learning with access to test queries or offline learning requiring high-quality annotated demonstrations.

    “Both methods also focus on online learning with access to test queries or offline learning requiring high-quality annotated demonstrations.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Wang et al., 2024e

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Sarch et al., 2024 source

      The Abstract (p. 1) says ICAL turns suboptimal, imperfect demonstrations into high-quality examples, contrary to the claim that it requires high-quality annotated demonstrations. The Overview (p. 4) says the learned examples are used during inference on unseen instructions and environments; it does not describe learning with access to test queries. (LLM, conf=0.85)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    On WebArena, SkillWeaver achieves higher average success rates and better or comparable performance across all domains compared to AutoEval, except in the Shopping environment.

    “When compared to AutoEval (Pan et al., 2024), which leverages an LLM-based reward model to guide inference-time exploration, SKILLWEAVER achieves higher average success rates and exhibits better or comparable performance across all domains, with the exception of the Shopping environment.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Pan et al., 2024 source
    • Rests on an assumption: WebArena is a good platform for demonstrating the efficacy of the agent.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Parametric training with exploration trajectories has been investigated using supervised fine-tuning and reinforcement learning-style optimization.

    “Parametric training with exploration trajectories has been investigated using various strategies, such as supervised fine-tuning (e.g., WebGUM (Furuta et al., 2024), Patel et al. (2024), and ScribeAgent (Shen et al., 2024)) and reinforcement learning-style optimization (e.g., ETO (Song et al., 2024a) and PAE (Zhou et al., 2024b)).”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Furuta et al., 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Song et al., 2024a source
    • cites Zhou et al., 2024b

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    On WebArena, integrating synthesized APIs yields a relative improvement in success rate of 39.8% on average for GPT-4o and 54.3% for GPT-4o-mini.

    “we observe a substantial relative improvement in success rate, 39.8% on average, for the baseline agent with GPT-4o and an even larger improvement of 54.3% with GPT-4o-mini across the evaluated websites.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: WebArena is a good platform for demonstrating the efficacy of the agent.

  • May combine two claims.

    Findingown data

    The agent achieves stable performance improvements across all 5 WebArena websites and 4 real-world live websites.

    “our agent is not only able to achieve stable performance improvements across all 5 WebArena websites but also in the 4 real-world live websites that are more complex and noisy.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Manual evaluation of agent trajectories on real-world websites is valid for determining success.
    • Rests on an assumption: WebArena is a good platform for demonstrating the efficacy of the agent.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The task 'request quote for multiple listings' requires repeated, sequential, successful searches for specific vehicles, followed by separate quote requests.

    “the task request quote for multiple listings requires repeated, sequential, successful searches for specific vehicles, followed by separate quote requests.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.

  • May combine two claims.

    Findingown data

    LLMs like GPT-4o are not robust enough at API calling, which hurts API-augmented agents.

    “LLMs like GPT-4o are still not robust enough at API calling, and that hurts API-augmented agents like ours, even with human-crafted APIs.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.

  • May combine two claims.

    Findingown data

    Two primary categories of failures in API use are failure to identify the appropriate API and generating wrong parameters.

    “We identify two primary categories of failures: (1) failure to identify the appropriate API and (2) generating wrong parameters.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    The agent found a bug in one of the APIs and was able to make recoveries.

    “We found a bug in one of the APIs, and were able to make the following recoveries.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    A skill that automatically performs the checkout process with the items currently in the cart was implemented on the shopping website.

    “Here is an skill that was implemented on the shopping website.It automatically performs the checkout process with the items currently in the cart.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    APIs are able to successfully automate multiple atomic actions and represent them as a single action with richer input parameters.

    “APIs are able to successfully automate multiple atomic actions and represent them as a single action with richer input parameters.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    On WebArena, adding skills improves average task success rate from 40.2% (baseline) to 56.2% (+ Skills), a relative improvement of 40%.

    “| Baseline | 65.0 | 11.7 | 62.5 | 11.1 | 40.2 |”
    • Also stated in: § D.2.3 Ceiling of Agent Capabilities · plain
      “| + Skills | 87.0 | 29.4 | 75.0 | 11.1 | 56.2 |”
    • Also stated in: § D.2.3 Ceiling of Agent Capabilities · plain
      “| ∆ | ↑ 34% ↑ 151% | ↑ 20% | ↑ 0% ↑ 40% | | |”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingreasoning

    In the following example, the agent attempts to learn how to add collaborators to a Gitlab repository, but is unable to because it does not have a source of real test input data.

    “In the following example, the agent attempts to learn how to add collaborators to a Gitlab repository, but is unable to because it does not have a source of real test input data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Adding 'if' statements to avoid atomic actions producing errors reduces the number of exceptions but does not improve the robustness of the API.

    “While this does reduce the number of exceptions, it does not improve the robustness of the API.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    For the task 'Browse recipes for gluten-free chocolate chip cookies that can be made without nuts,' the agent incorrectly uses the parameter search_recipes_by_ingredients(page, 'chocolate chip, -nuts').

    “the agent incorrectly uses the parameter search recipes by ingredients(page, 'chocolate chip, -nuts')”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: For the task 'Browse recipes for gluten-free chocolate chip cookies that can be made without nuts,' the correct parameter is search_recipes_by_ingredients(page, 'chocolate chip without nuts').

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    SkillWeaver comprises a three-stage pipeline: systematic exploration of website environments to identify potential skills, practicing and converting these skills into robust, reusable APIs, and testing and debugging to ensure API reliability during inference.

    “Specifically, SKILLWEAVER comprises a three-stage pipeline (Figure 1): (1) systematic exploration of website environments to identify potential skills, (2) practicing and converting these skills into robust, reusable APIs, and (3) testing and debugging to ensure API reliability during inference.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.
    • Rests on an assumption: LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.
    • Rests on an assumption: A curriculum progression from simple to complex skill compositions improves accuracy because modules perform better on simpler tasks requiring fewer steps.
    • Rests on an assumption: Static analysis of generated Python functions detects common generation mistakes.
    • Rests on an assumption: Unit testing and debugging synthesized APIs ensures their reliability.

  • May combine two claims.

    Interpretationreasoning

    The task 'request quote for multiple listings' demands both long-range planning and back-tracing capabilities, which present significant challenges for existing agents.

    “This process demands both long-range planning and back-tracing capabilities, which present significant challenges for existing agents.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.

  • May combine two claims. The quote only partly states this claim.

    Interpretationown data

    The performance of synthesized APIs is comparable to or even superior to manually crafted APIs found in official documentation.

    “This result suggests that our exploration process is capable of generating APIs with quality that is comparable to or even superior to manually crafted APIs found in official documentation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Human-crafted APIs extracted from official documentation are a fair comparison baseline for synthesized APIs.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Skill Honing validates APIs through unit testing and debugging to ensure reliability.

    “Additionally, the subsequent Skill Honing module (Stage III) validates APIs through unit testing and debugging to ensure reliability.”
    • Also stated in: § 2.3 Stage III: Skill Honing · plain
      “we integrate a stage to test and debug synthesized APIs.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Unit testing and debugging synthesized APIs ensures their reliability.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The reason for the Shopping exception is the inherent requirements for more extensive interaction with such websites for partially observable information, like dynamic product search results and product details.

    “The reason behind this exception is the inherent requirements for more extensive interaction with such websites for partially observable information, like dynamic product search results and product details.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.

  • Probably not a substantial claim: may not pass S4 (based).

    Study design and conductnone given

    Manual evaluation of agent trajectories on real-world websites ensures validity of evaluation.

    “To ensure the validity of evaluation, we manually evaluate the success of agent trajectories.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Manual evaluation of agent trajectories on real-world websites is valid for determining success.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Static analysis of generated functions detects common generation mistakes, and if any are detected, the model regenerates the function.

    “We statically analyze the function for common generation mistakes, as described in Appendix § A.6, and if any are detected, we prompt the model to generate the function again.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Static analysis of generated Python functions detects common generation mistakes.

  • May combine two claims.

    Premisecitation

    Existing methods rely on a set of existing training examples to determine tool correctness, or user input queries to bootstrap the tool-making system.

    “However, existing methods rely on a set of existing training examples to determine tool correctness (Yuan et al., 2024a), or user input queries to bootstrap the tool-making system (Wang et al., 2024d).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Yuan et al., 2024a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Wang et al., 2024d

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Non-parametric methods typically transform exploration trajectories into reusable APIs or workflows that can be composed into more complex skills.

    “Non-parametric methods typically transform exploration trajectories into reusable APIs or workflows that can be composed into more complex skills, like Voyager (Wang et al., 2023a) an LLM-based lifelong learning agent that continually explores the Minecraft environment.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Wang et al., 2023a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    During online evaluation, the authors manually validated the safety of all actions before execution.

    “During online evaluation, we noticed the possibility of these web agents generating harmful actions on the web, and we manually validated the safety of all the actions before execution.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Manual validation of all actions before execution ensures no harmful actions occur during online evaluation.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    The agent needs better parameter selection logic to enhance performance.

    “This indicates that the agent needs better parameter selection logic to enhance performance.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    SkillWeaver focuses on enabling web agents to continuously expand their skill repertoire by autonomously exploring website environments and distilling successful trajectories into reusable skills represented as Python APIs.

    “Our work focuses on enabling web agents to continuously expand their skill repertoire by autonomously exploring website environments and distilling successful trajectories into reusable skills represented as Python APIs.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Website environments contain numerous underlying functionalities that require systematic exploration.
    • Rests on an assumption: LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationown data

    Some failure cases seem to be caused by gaps in LLM reasoning capabilities for certain situations.

    “We observe some failure cases, which seem to be caused by gaps in LLM reasoning capabilities for certain situations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Screenshot and HTML files consume approximately 0.3 GB per trajectory with an average of 7.3 actions in Mind2Web, making them hard to transfer on the fly.

    “For example, screenshot and HTML files consume approximately 0.3 GB per trajectory with an average of 7.3 actions in Mind2Web(Deng et al., 2023), making them hard to transfer on the fly.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Deng et al., 2023

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    To ensure robustness, the synthesized API undergoes testing with automatically generated test cases and debugging within the Skill Honing module (Stage III).

    “To ensure robustness, the synthesized API undergoes testing with automatically generated test cases and debugging within the Skill Honing module (Stage III).”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The automatically generated test cases adequately test the synthesized API's robustness.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Some skills require interacting with data stored on the website, and their success is conditioned not on whether the UI is accurately interacted with but rather whether the function inputs are realistic.

    “Some skills require interacting with data stored on the website, and their success is conditioned not on whether the UI is accurately interacted with but rather whether the function inputs are realistic.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    In case the agent is being used on a live website, the authors add safety instructions that guardrail the agent away from actions with potential side effects, such as creating accounts or interacting with human users.

    “In case the agent is being used on a live website, we add safety instructions which guardrail the agent away from actions that could potentially have side effects or that would otherwise be disadvantageous to explore, such as creating accounts or interacting with human users.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The safety instructions effectively prevent the agent from taking harmful actions on live websites.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data

    The agent's policy needs improvement to better utilize generated APIs.

    “suggesting the need for improving the agent's policy to better utilize generated APIs”
    Quote in the paperNot a substantial claim
33 of 33

  • The paper relies on this without support.

    Website environments contain numerous underlying functionalities that require systematic exploration.

    “Website environments encompass numerous underlying functionalities that require systematic exploration.”

  • The paper relies on this without support.

    LLMs possess internet-scale knowledge and HTML interpretation capabilities sufficient to propose novel, reusable skills for website environments.

    “Similarly, we leverage LLMs as an automatic curriculum for discovering website functionalities, leveraging their internet-scale knowledge and HTML interpretation capabilities.”

  • The paper relies on this without support.

    Accessibility trees provide comprehensive structural and functional information for understanding websites.

    “Accessibility trees are particularly useful for understanding less familiar websites, such as CMS in WebArena, because they provide comprehensive structural and functional information.”

  • The paper relies on this without support.

    A curriculum progression from simple to complex skill compositions improves accuracy because modules perform better on simpler tasks requiring fewer steps.

    “We address this challenge through curriculum progressionadvancing from simple to complex skill compositions-which improves accuracy since modules perform better on simpler tasks requiring fewer steps.”

  • The paper relies on this without support.

    Static analysis of generated Python functions detects common generation mistakes.

    “We statically analyze the function for common generation mistakes, as described in Appendix § A.6, and if any are detected, we prompt the model to generate the function again.”

  • The paper relies on this without support.

    Unit testing and debugging synthesized APIs ensures their reliability.

    “Additionally, the subsequent Skill Honing module (Stage III) validates APIs through unit testing and debugging to ensure reliability.”

  • The paper relies on this without support.

    Manual evaluation of agent trajectories on real-world websites is valid for determining success.

    “To ensure the validity of evaluation, we manually evaluate the success of agent trajectories.”

  • The paper relies on this without support.

    WebArena is a good platform for demonstrating the efficacy of the agent.

    “WebArena provides a stable testing environment with reliable automatic evaluation, making it a good platform for demonstrating the efficacy of our agent and analyzing various aspects of our approach.”

  • The paper relies on this without support.

    Human-crafted APIs extracted from official documentation are a fair comparison baseline for synthesized APIs.

    “We further leverage these human-crafted official APIs to compare them against the APIs synthesized by SKILLWEAVER, demonstrating the quality of the generated APIs.”

  • The paper relies on this without support.

    Natural language-based routines pose challenges for formal verification and precise composition into new workflows.

    “However, its natural language-based routines pose challenges for formal verification and precise composition into new workflows.”

  • The paper relies on this without support.

    Traditional trajectory-based approaches struggle to explicitly abstract reusable procedural knowledge, resulting in heavy training demands and limited generalization to new websites and tasks.

    “Although these trajectory-based approaches can be effective, they struggle to explicitly abstract reusable procedural knowledge, resulting in heavy training demands and limited generalization to new websites and tasks.”

  • The paper relies on this without support.

    The automatically generated test cases adequately test the synthesized API's robustness.

    “To ensure robustness, the synthesized API undergoes testing with automatically generated test cases and debugging within the Skill Honing module (Stage III).”

  • The paper relies on this without support.

    The safety instructions effectively prevent the agent from taking harmful actions on live websites.

    “In case the agent is being used on a live website, we add safety instructions which guardrail the agent away from actions that could potentially have side effects or that would otherwise be disadvantageous to explore, such as creating accounts or interacting with human users.”

  • The paper relies on this without support.

    Manual validation of all actions before execution ensures no harmful actions occur during online evaluation.

    “During online evaluation, we noticed the possibility of these web agents generating harmful actions on the web, and we manually validated the safety of all the actions before execution.”

  • The paper relies on this without support.

    For the task 'Save a hamburger recipe,' the correct first API call is search_recipes_by_cuisine_type('hamburger').

    “the agent should first call search recipes by cuisine type('hamburger') to obtain recipe details and then select the print option to save the recipes.”

  • The paper relies on this without support.

    For the task 'Browse recipes for gluten-free chocolate chip cookies that can be made without nuts,' the correct parameter is search_recipes_by_ingredients(page, 'chocolate chip without nuts').

    “the correct parameter should be search recipes by ingredients(page, 'chocolate chip without nuts')”
16 of 16

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Agents can refine tool-use capabilities through demonstrations and environmental feedback across diverse domains such as mobile GUIs, web navigation, and software engineering.

    “Recent advancements empower agents to autonomously create specialized tools for problemsolving (Qian et al., 2023; Yuan et al., 2024a), and to refine tool-use capabilities through demonstrations and environmental feedback across diverse domains such as mobile GUIs, web navigation, and software engineering (Fang et al., 2025d; Zheng et al., 2025a; Bouzenia et al., 2024).”

    From Memory in the Age of AI Agents (2025) · cites this paper as Zheng et al., 2025a

    The Self-Improvement and Stage III: Skill Honing sections describe a web agent that learns from exploration trajectories and tests and debugs its synthesized skills. This supports the web-navigation and feedback-based refinement portion of the claim. The Introduction says the method does not require annotated demonstrations, and this paper does not establish refinement in mobile GUIs or software engineering. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Agents refine tool-use capabilities through demonstrations and environmental feedback across diverse domains such as mobile GUIs, web navigation, and software engineering.

    “Recent advancements empower agents to autonomously create specialized tools for problemsolving (Qian et al., 2023; Yuan et al., 2024a), and to refine tool-use capabilities through demonstrations and environmental feedback across diverse domains such as mobile GUIs, web navigation, and software engineering (Fang et al., 2025d; Zheng et al., 2025a; Bouzenia et al., 2024).”

    From Memory in the Age of AI Agents (2025) · cites this paper as Zheng et al., 2025a

    The Self-Improvement section and §3.4 describe web agents refining reusable skills through autonomous exploration, practice, and testing, supporting the web-navigation and environmental-feedback portion. The Introduction and Tool Use and Synthesis section distinguish this approach from learning through annotated demonstrations; this paper does not support that part of the attribution. Mobile GUIs and software engineering appear to be covered by the sentence’s other citations, not this one. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

  • Skill induction reads the supply side, what an agent did while acting, and so observes a routine only after an agent has already performed it at full cost, and only for tasks it can complete.

    “All read the supply side, what an agent did while acting, and so observe a routine only after an agent has already performed it at full cost, and only for tasks it can complete.”

    From Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay (2026) · cites this paper as [38]

    The API Synthesis section says SkillWeaver turns an agent’s successful action trajectories into reusable APIs, supporting the point that a learned routine follows a completed attempt. But Stage I: Skill Proposal says it proposes skills before practice, and §3.4 Exploration says the agent can use existing APIs while practicing. The source therefore does not support the unqualified claims that it observes routines only after performing them at full cost or that skill discovery is limited to tasks it can already complete. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • In prior self-improvement systems such as EvoSkill, GEPA, and SkillWeaver, the meta-level logic is hardcoded in advance and shared by every branch throughout the run.

    “These systems, however, evolve only what the agent does, not how it evolves : the artifact under optimization changes while the operator that optimizes it stays fixed. In the vocabulary of selfimproving machines (Good, 1965; Schmidhuber, 2006), they are self-improving but stop short of being recursively self-improving. The meta-level logic is hardcoded in advance and shared by every branch through”

    From MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution (2026) · cites this paper as Zheng et al., 2025

    Section 2, “SkillWeaver,” describes a three-module exploration pipeline that synthesizes and refines skills as APIs, supporting the distinction between changing skills and a specified improvement process. The paper does not say that its meta-level logic is hardcoded or shared by every branch throughout a run; it does not establish the claim’s branch-level qualifier. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

  • Few methods generate tasks to build memory instead of updating model weights.

    “Few methods generate tasks to build memory instead: Voyager (Wang et al., 2024) proposes tasks through an automatic curriculum, SkillWeaver (Zheng et al., 2025) practices self-proposed skills on websites, and PREPING (Choi et al., 2026), closest to our setting, builds a playbook from self-proposed tasks, updating it from each task's feasibility and completion scores.”

    From DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks (2026) · cites this paper as Zheng et al., 2025

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 9

  1. OpenReview.net, 2024. URL https://openreview.net/forum?id=O4cHTxW9BS.
  2. Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. Guiding pretraining in reinforcement learning with large language models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pp. 8657-8677. PMLR, 2023. URL https: //proceedings.mlr.press/v202/du23f.html.
  3. Hiroki Furuta, Kuang-Huei Lee, Ofir Nachum, Yutaka Matsuo, Aleksandra Faust, Shix- iang Shane Gu, and Izzeddin Gur. Multimodal web navigation with instruction- finetuned foundation models. In The Twelfth International Conference on Learning Rep- resentations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. URL https://openreview.net/forum?id=efFmBWioSc.
  4. Wei Li, William E. Bishop, Alice Li, Christopher Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on UI control agents. In Amir Globersons, Lester Mackey, Danielle Belgrave, An- gela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang (eds.), Advances in Neural Information Processing Systems 38: Annual Conference on Neural Infor- mation Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 -15, 2024, 2024. URL http://papers.nips.cc/paper files/paper/2024/hash/ a79f3ef3b445fd4659f44648f7ea8ffd-Abstract-Datasets and Benchmarks Track.html.
  5. Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (eds.), Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 -16, 2023, 2023. URL http://papers.nips.cc/paper files/paper/ 2023/hash/871ed095b734818cfba48db6aeb25a62-Abstract-Conference.html.
  6. Haiming Wang, Huajian Xin, Chuanyang Zheng, Zhengying Liu, Qingxing Cao, Yinya Huang, Jing Xiong, Han Shi, Enze Xie, Jian Yin, Zhenguo Li, and Xiaodan Liang. LEGO- prover: Neural theorem proving with growing libraries. In The Twelfth International Conference on Learning Representations, 2024a. URL https://openreview.net/forum?id= 3f5PALef5B.
  7. Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh (eds.), Ad- vances in Neural Information Processing Systems 35: Annual Conference on Neural Infor- mation Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. URL http://papers.nips.cc/paper files/paper/2022/hash/ 82ad13ec01f9fe44c01cb91814fd7b8c-Abstract-Conference.html.
  8. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. URL https://openreview.net/forum?id=WE vluYUL-X.
  9. Keen You, Haotian Zhang, Eldon Schoop, Floris Weers, Amanda Swearngin, Jeffrey Nichols, Yinfei Yang, and Zhe Gan. Ferret-ui: Grounded mobile UI understanding with multimodal llms. In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G ül Varol (eds.), Computer Vision -ECCV 2024 -18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXIV, volume 15122 of Lecture Notes in Computer Science, pp. 240-255. Springer, 2024. doi: 10.1007/978-3-031-73039-9\ 14. URL https://doi.org/10.1007/978-3-031-73039-9 14. · DOI · Paper page

Cited by

Sorted by citations

Authors
Zheng, B., Fatemi, M. Y., Jin, X., Wang, Z. Z., Gandhi, A., Song, Y., Gu, Y., Srinivasa, J., Liu, G., Neubig, G., & Su, Y.
Affiliations
  • The Ohio State University (US)
  • University of Virginia (US)
  • Carnegie Mellon University (US)
ORCID
Venue
arXiv
Publisher
arXiv
Published
2025
ISSN
2331-8422 (ISSN-L)
DOI
10.48550/arxiv.2504.07079
arXiv
2504.07079
OpenAlex
W4417248776
Cited by
0
References
0
Type
posted content
Language
en
Open access
Licence
Locations
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
Cisco Systems
Topics
  • Physical Sciences › Computer Science › Information Systems › Web Data Mining and Analysis
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
Keywords
skill discovery, skill composition, procedural knowledge, human development, skill transfer, WebArena

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-24.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-24.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review