Skip to content

PolySkill: Learning Generalizable Skills Through Polymorphic Abstraction For Continual Learning

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Polymorphic skill abstraction for continual learning in LLM agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

PolySkill applies object-oriented polymorphism to web-agent skill learning, separating a skill's abstract interface (e.g., AbstractShoppingSite.search_product()) from its site-specific concrete implementations. This decoupling enables skills to transfer across websites in the same functional domain without re-specialization. The framework achieves up to 9.4% higher success rates on Mind2Web and 13.9% on unseen websites, with 1.7Ă— better skill reuse.

Problem

Existing skill-induction methods (ASI, SkillWeaver) generate over-specialized skills tied to a single website's UI, yielding skill reusability below 18% on unseen websites. There is no principled mechanism to separate a skill's semantic intent from its concrete implementation, preventing cross-website and cross-domain transfer. Prior evaluation also relied solely on task success rate, obscuring whether agents actually reuse skills.

Method

PolySkill organizes learned skills into a domain-driven class hierarchy inspired by OOP polymorphism. For each functional website category (e.g., shopping), the agent first induces an abstract class (AbstractShoppingSite) defining method signatures (search_product(query), add_to_cart(item_id, quantity), checkout()). When visiting a new site (e.g., walmart.com), the agent instantiates a concrete subclass (WalmartWebsite) that inherits the abstract interface and fills in site-specific implementations. Compositional skills (e.g., purchase_item = find_and_add_to_cart + checkout) are defined once at the abstract level and inherited automatically.

The induction pipeline (Algorithm 1) follows a verify-then-induct loop: execute task → LM judge verifies trajectory → LM induces hierarchical skill → add to dynamic library \(K_t\). The agent policy is \(\pi_L(a_t | o_t, M_t, K_t)\) over the expanded action space \(A_t = A_p \cup K_t\), optimizing an efficiency-aware reward \(\max_{\pi_L, K} \mathbb{E}_{q \sim Q}[g(\tau, q) - \gamma|\tau|]\).

The framework also supports a task-free continual learning setting where the agent self-proposes tasks guided by which abstract method signatures remain unimplemented, providing structured exploration.

Key Contributions

  • Polymorphic skill representation: abstract class + multiple concrete subclass implementations, enabling cross-website generalization without re-learning compositional logic
  • Three new evaluation metrics: Skill Reusability (fraction of induced skills reused on new tasks), Task Coverage (fraction of test tasks that invoke at least one skill), Skill Compositionality (rate at which existing skills are chained to form new ones)
  • Empirical demonstration of cross-website skill transfer: 31% reuse rate for PolySkill vs. <9% for ASI and <3% for SkillWeaver on unseen websites
  • First comprehensive skill-induction evaluation on open-source agentic models (Qwen3-Coder-480B-A35B, GLM-4.5) alongside proprietary models
  • Task-free continual learning: abstract class schemas act as structured exploration priors, improving self-proposed task quality

Results

  • Mind2Web Cross-Domain (GPT-4.1): PolySkill (+Update) reaches ~63–64% vs. ~62% for ASI (+Update)
  • Mind2Web Cross-task (GPT-4.1): PolySkill improves from 52–53% (ASI) to 55–56%; online variant to ~63% vs. 59% (ASI +Update)
  • Mind2Web (Qwen3-Coder): Cross-task 47.5% vs. 41.5% (ASI); Cross-Domain 39.9% vs. 35.2% (ASI)
  • Unseen websites: up to 13.9% success rate improvement over baselines
  • Skill Reusability on unseen websites: 31% (PolySkill) vs. <9% (ASI) and <3% (SkillWeaver)
  • Step reduction: over 20% fewer steps; inverse correlation between skill reusability and steps confirmed across all three compared methods
  • Skill reuse on seen websites: 1.7Ă— improvement over prior methods

Limitations

  • Evaluation restricted to web navigation (Mind2Web, WebArena); generalization to other agent domains (e.g., desktop GUIs, robotics) is asserted but not empirically demonstrated
  • Requires the agent to correctly identify the functional domain of a new website to retrieve the right abstract class; mismatch degrades transfer
  • The LM judge (GPT-4.1 based) used for trajectory verification achieves only 85% agreement with human judgment, introducing noise in skill induction
  • Task-free continual learning results are preliminary; structured exploration gains are shown qualitatively rather than with full ablation
  • Abstract class design is currently hand-scoped by domain category; automatic domain taxonomy induction is not addressed

Relevance to Agentic AI / LLM Agents

PolySkill directly advances the skill-induction sub-field of agentic AI by solving the generalization bottleneck that has limited prior systems (Voyager, AWM, ASI, SkillWeaver) to single-environment specialization. The polymorphic abstraction principle—separating interface from implementation—is a transferable design pattern for any LLM agent operating across structurally similar but surface-distinct environments, not just web tasks. The task-free continual learning setting is particularly relevant for the broader goal of self-improving agents that autonomously expand their skill libraries. New metrics (Reusability, Coverage, Compositionality) address a measurement gap in the field and provide a more rigorous scaffold for evaluating future skill-learning systems.