PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction

• PolySkill improves task success rate, skill reusability, and compositional capabilities by 1.3–1.8× compared with existing methods.Figure 1 presents the overall comparison across these capability dimensions.Source: Figure 1 · p. 2, §1 · p. 3
• Existing skill-induction methods optimize familiar websites and therefore leave cross-website generalization under-explored.This creates a tension between skill specificity and generalizability and motivates measuring transfer and reuse directly.Source: §1 · pp. 1–2
• PolySkill separates abstract skill goals from website-specific implementations, enabling compositional skills that are less tied to individual interfaces.Abstract operations can be chained across concrete website subclasses, such as search, add-to-cart, and checkout.Source: §1 · p. 2
• 31% reuse on unseen websites contrasts with below 18% for prior methods, showing that PolySkill directly improves measured skill transfer.The evaluation adds Skill Reusability and Task Coverage because task success alone cannot distinguish reuse from solving tasks from scratch.Source: §1 · p. 2
• In task-free continual learning, PolySkill better guides exploration and supports skill induction without predefined tasks.The setting tests whether agents can explore multiple websites and learn skills autonomously.Source: §1 · p. 3

• Concrete descriptions, action traces, and website-specific programs hinder reuse across environments with similar functions.These representations lack a mechanism for separating semantic intent from implementation.Source: §2 · p. 3
• PolySkill represents each skill through an abstract interface linked to multiple interchangeable concrete implementations.This hybrid representation preserves programmatic execution while adding flexibility across web pages.Source: §2 · p. 3
• Continual learning requires compositional generalization, but existing web-agent skills are difficult to modify or combine systematically.Polymorphism addresses variation between implementations while maintaining a stable interface.Source: §2 · p. 3

3.2 THE POLYSKILL FRAMEWORK

• Existing methods produce unstable, over-specialized skills that transfer poorly to unseen websites, with Skill Reusability below 9% for ASI and below 3% for SkillWeaver.The reported instability includes increasingly complex and specific self-proposed tasks for SkillWeaver with Claude-3.7-Sonnet.Source: §3.2 · pp. 4–5, Figure 2 · p. 4
• PolySkill separates abstract skill goals from concrete website implementations through domain-level classes and site-specific subclasses.For example, AbstractShoppingSite provides shared signatures while AmazonWebsite implements them concretely.Source: Table 1 · p. 5, §3.2 · p. 5, §3.3 · p. 5

3.3 SKILL INDUCTION PROCESS

• The induction pipeline executes tasks, verifies successful trajectories, induces hierarchical skills, and adds them to the dynamic library only after validation.PolySkill extends the ASI verification pipeline by registering abstract signatures before defining site-specific implementations.Source: §3.3 · pp. 5–6

3.4 EVALUATION SETUP

• A known domain abstraction guides exploration on new websites toward implementing essential skills such as search and add-to-cart, accelerating structured skill acquisition.The framework can reuse an abstract shopping blueprint when moving from Amazon to Walmart, then create a Walmart-specific implementation.Source: §3.3 · p. 6, §3.4 · p. 6
• PolySkill is evaluated in task-defined Mind2Web and WebArena benchmarks, plus task-free continual learning where agents propose goals and induce skills through exploration.The task-free setting tests whether learned abstract classes provide a schema for deciding which skills are worth discovering.Source: §3.4 · p. 6

3.5 EVALUATION METRICS

• Evaluation adds Skill Reusability, Task Coverage, and Skill Compositionality to Task Success Rate and Number of Steps, measuring utility, adaptability, and reuse of learned building blocks.Skill Compositionality captures how often existing skills are reused to construct more complex tasks.Source: §3.5 · pp. 6–7
• Mind2Web results show PolySkill consistently outperforming ASI, with PolySkill (+Update) achieving state-of-the-art performance across models and increasing generalization settings.The figure compares Cross-task, Cross-Website, and Cross-Domain success rates, with the largest gains in Cross-Domain.Source: Figure 3 · p. 7

4.1 STANDARD BENCHMARK EVALUATION

• PolySkill improves Mind2Web Cross-task and Cross-Domain performance over ASI across GPT-4.1 and Qwen3-Coder, especially with online updating.GPT-4.1 reaches about 63% online Cross-task accuracy versus 59% for ASI + Update, while Qwen3-Coder reaches 47.5% Cross-task and 39.9% Cross-Domain.Source: §4.1 · p. 8

4.2 ANALYSIS

• As skill reusability rises from 0% to over 20%, average task steps fall from approximately 6.1 to 3.3–4.4, linking reuse with efficiency.PolySkill reaches 20.4% skill reusability by task 180, while ASI shows the largest step reduction despite lower utilization.Source: §4.2 · p. 8, Figure 4 · p. 8
• PolySkill maintains a smaller static skill library while improving accuracy, reaching 55.4% Cross-task accuracy with 43 skills versus ASI’s 52.3% and 50 skills.The results support the use of polymorphic skills across interface variations rather than separate specialized routines.Source: §4.2 · pp. 8–9, Table 2 · p. 9
• PolySkill + Update reaches 63.2% Cross-task accuracy with 47 skills, compared with 66 skills for ASI + Update.This combines stronger performance with a smaller skill library in the online setting.Source: §4.2 · p. 9
• During continual learning, PolySkill transfers to Amazon and Target while preserving original WebArena performance, ending 4.9% above ASI.ASI’s original-task performance degrades after adapting to new sites, whereas PolySkill remains stable.Source: §4.2 · p. 9

4.3 FROM SPECIALIST TO EXPLORER: SKILL LEARNING IN AN EXPLORATIVE SETTING

• Self-guided exploration achieves 43.1% success on held-out WebArena Shopping and leads both GitLab at 66.2% and GitHub at 84.0%.The autonomous setting outperforms single-domain specialists and handcrafted curriculum approaches in cross-site transfer.Source: §4.3 · p. 10

5 CONCLUSION

• 73% of learned skills transferred to unseen websites, compared with <31% for prior methods.Source: §5 · p. 10
• PolySkill separates a skill’s high-level intent from its website-specific implementation, enabling reuse across diverse environments.Source: §5 · p. 10

5.1 FUTURE WORK

• Future work includes handling highly dynamic websites, sharing skills between agents, and incorporating human feedback.Source: §5 · p. 10
• The framework could move from manual domain labels to autonomous skill clustering based on functional similarity.Source: §5.1 · p. 10

• Experiments use an 8-H100 GPU cluster, official APIs for proprietary LLMs, and BrowserGym for all experimental interactions.Source: REPRODUCIBILITY STATEMENT · p. 11

ETHICS STATEMENT

• Web agents could misuse websites or violate terms of use and copyright, so access may be removed at an owner’s request.Source: ETHICS STATEMENT · p. 11
• Agent deployment requires checks for hidden bias and safeguards because agents can take actions in the world.Source: ETHICS STATEMENT · p. 11
• The systems are research prototypes intended for research use and may not be ready for real-world or high-stakes deployment.Source: ETHICS STATEMENT · p. 11

A LARGE LANGUAGE MODEL USAGE

• The authors used LLMs only for language refinement and literature searching, not ideation, methodology, design, experiments, or analysis.Source: A LARGE LANGUAGE MODEL USAGE · p. 15

B EXTENDED RELATED WORKS

• Prior skill representations include textual descriptions, action traces, and programs, but programmatic skills remain tied to specific contexts.Source: B EXTENDED RELATED WORKS · pp. 15–16
• Broader agent-memory research structures episodic experiences, organizes working memory hierarchically, or maintains persistent external stores.Source: B EXTENDED RELATED WORKS · p. 16

C LIMITATIONS

• PolySkill is shown to work with latest agentic models for continual learning, while smaller models remain a proposed direction for further reinforcement-learning training.Source: C LIMITATIONS · p. 16

D SKILL EXAMPLES

• Examples of ASI and SkillWeaver skills are provided in Table 4, while PolySkill examples are provided in Table 1.Source: D SKILL EXAMPLES · p. 16

E.1 FORMAL DEFINITIONS OF EVALUATION METRICS

• Task Success Rate is the fraction of evaluation tasks completed successfully, using an indicator for each task outcome.Source: §E.1 · p. 17
• Number of Steps averages action counts only over successful trajectories to measure efficiency.Source: §E.1 · p. 17
• Skill Reusability measures the fraction of learned library skills used at least once during evaluation.Source: §E.1 · p. 17
• Skill Adoption Rate measures the fraction of test trajectories invoking at least one induced skill.Source: §E.1 · p. 17
• Skill Compositionality measures the average number of previously learned skills reused inside each newly created skill.Source: §E.1 · p. 17

Mind2Web

• Mind2Web contains over 2,350 real-world tasks across 137 websites and 31 domains, testing agents’ ability to follow instructions through web actions.Each task pairs a natural-language instruction with a specific website and evaluates successful action sequences.Source: §E.3 · p. 17

Action space

• BrowserGym supplies the base primitive action space used by the evaluation environment.Source: Table 5 · p. 18

WebArena

• WebArena provides 811 compositional tasks across five self-hosted applications, including e-commerce, forums, software development, maps, and administration.Many tasks require navigating and synthesizing information across multiple applications.Source: §E.3 · p. 18

E.4 REAL-WORLD WEBSITE TASKS

• The Amazon tasks cover constrained product discovery using brands, attributes, prices, ratings, shipping eligibility, and sorting requirements.Examples include Herman Miller chairs, gaming monitors, air fryers, laptops, robotic vacuums, televisions, and other products.Source: §E.4 · pp. 18–19
• The benchmark also tests product-page lookup and review understanding, including specifications, compatibility, nutrition, battery life, criticism, and topic-specific customer comments.Tasks ask agents to inspect item details, Q&A sections, and reviews for products such as headphones, appliances, books, electronics, food, and clothing.Source: §E.4 · p. 19
• State-changing shopping tasks require adding specified products and quantities to carts, wish lists, custom lists, registries, or checkout flows without necessarily placing orders.Examples include cart quantity updates, private wish-list creation, comments on saved products, baby registries, and checkout or pickup selection.Source: §E.4 · pp. 19, 21
• Target tasks extend these capabilities to retail categories, deals, store services, policies, pickup, and account-related navigation, while the GitHub set contains 75 navigational or non-code tasks performed with dummy accounts.GitHub examples include editing profiles, creating or configuring repositories, starring or forking repositories, inspecting commits and contributors, and managing issues.Source: §E.4 · pp. 20–22
• The real-world task set covers GitHub actions including commenting on issues and pull requests, creating branches and pull requests, managing collaborators and milestones, searching repositories, and changing repository settings.Examples include enabling GitHub Pages, protecting a branch, creating organizations, viewing releases, and following users.Source: §E.4 · pp. 22–23

E.5 PROMPTS

• The skill-induction prompt asks an agent to summarize reusable functions from action trajectories and rewrite examples using functions with 2–5 task-oriented steps, common arguments, documentation, and no example-specific information.Functions should have at least three actions, moderate complexity, and may be generated zero, one, or multiple at a time.Source: §E.5 · p. 24
• The WebJudge prompts extract explicit task key points, assess whether screenshots contain necessary completion evidence, and judge success using action histories, relevant snapshots, filters, submissions, and visible progress.The outcome judge returns success or failure after checking exact requirements, applied filters, visible results, and whether the agent avoided non-progressing loops.Source: §E.5 · pp. 24–25

Complete results

• Complete benchmark results are provided in Tables 6 and 8.Source: F COMPLETE RESULTS · p. 26

Overall comparison

• PolySkill achieves the highest average WebArena success rate across all four evaluated language models.The largest gains over ASI and SkillWeaver occur with GPT-4.1 and Claude-3.7-Sonnet.Source: Figure 6 · p. 26
• Figure 6 compares PolySkill and baselines across WebArena website categories, with the Average group representing the primary overall result.Source: Figure 6 · p. 26

Measurement setup

• The analysis measures average input-plus-output tokens per Mind2Web cross-task instance using GPT-4.1 and separates inference from verification and skill-induction costs.Source: §F.1 · pp. 26–27

Overhead

• PolySkill incurs higher initial total token usage because Verification and Skill Induction add overhead to the no-skill baseline.Source: §F.1 · p. 27

Inference cost

• 10.2% lower Action Generation token consumption than the Baseline, from ∼55.4k to ∼49.7k tokens, and 7.7% lower than ASI.The reduction is attributed to solving tasks in fewer steps with higher precision.Source: §F.1 · p. 27

Amortized efficiency

• Verification and Skill Induction costs are primarily one-time or amortized after a website’s concrete implementation is stored and reused.As task volume increases, average cost per task converges toward the lower Action Generation cost.Source: §F.1 · p. 27

Benchmark evaluation

• WebArena performance is compared across website categories using success rates, with best and second-best results marked in the table.Some experiments are identified as pending completion.Source: Table 8 · p. 28

Skill transfer

• Self-guided exploration achieves the highest success rate on the held-out GitLab benchmark when learning from GitHub and GitLab concurrently.The comparison concerns transfer from a higher-performing domain to the more complex GitLab setting.Source: Table 9 · p. 28

Token efficiency

• PolySkill significantly reduces token usage during Mind2Web action generation despite the initial overhead of skill induction.The reduction lowers long-term costs per task instance.Source: Table 10 · p. 29

Task-free continual learning

• Task-free PolySkill replaces a predefined task sequence with iterative task proposal, execution, judging, skill induction, and observation updates.At each exploration step, the policy proposes a task from the current observation and skill library, then adds a skill only after successful judgment.Source: G EXTENDED METHODOLOGY · p. 29