Research map / Multi-agent systems and games / LLM agents

Beyond Domains: Reusing Web Skills via Transferable Interaction Patterns

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SkillMigrator, a web agent that stores induced skills as transferable interaction patterns (skill plus layout sketch) and retrieves them globally by layout and text similarity to reuse skills across websites and domains, cutting LLM action count on Mind2Web and WebArena.

Problem

LLM web agents make many policy LLM calls per task; existing web-skill libraries are retrieved by instruction similarity or site metadata, giving low skill reuse on held-out sites and domains.

Approach

Index each induced skill by a structural accessibility-tree sketch of its source page (a transferable interaction pattern), retrieve globally by combining layout tree-edit-distance with sentence-embedding text similarity, gate on a threshold, then bind slots to live element refs via Hungarian matching so skills replay without LLM calls.

Key results

• On WebArena, SkillMigrator reaches 45.7% average SR with 5.4 average LLM actions versus PolySkill's 49.3% SR and 5.9 actions (about 8.5% fewer actions).
• On Mind2Web cross-domain with +Update, SkillMigrator gets 63.0% SR at 6.2 actions versus PolySkill (+Update) 63.4% SR at 6.9 actions.
• Removing the layout signal (text only) drops WebArena SR from 45.7% to 39.2%, close to the ReAct baseline of 38.5%.

Surprising findings

• SkillMigrator has lower SR than PolySkill on WebArena (45.7 vs 49.3) and the paper reports it as competitive rather than best on success.
• Disabling the confidence gate lowers LLM action count further (5.4 to 4.7 on WebArena) at small SR cost.

Limitations

• Layout-based retrieval assumes pages with similar accessibility-tree structure share interaction semantics, which fails on visually isomorphic pages with different control flow.
• Operation templates are mined from observed action shapes, so unseen interaction types (drag-and-drop, modal dialogs) fall back to react mode.
• Text-only observations; multimodal cues are not used.

Experimental setup

Datasets: Mind2Web WebArena

Benchmarks: Mind2Web WebArena

Metrics: average LLM-action count on successful trajectories task success rate skill reuse rate library size

Baselines: ReAct SkillWeaver ASI PolySkill

Builds on: ASI PolySkill SkillWeaver ReAct AWM

Closest papers by meaning

• How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.88)
• From History to State: Constant-Context Skill Learning for LLM Agents (NeurIPS 2026 · similarity 0.87)
• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.87)
• RealWebAssist: A Benchmark for Long-Horizon Web Assistance with Real-World Users (AAAI 2026 · similarity 0.87)
• WALT: Web Agents that Learn Tools (ICLR 2026 · similarity 0.87)
• SkillOS: Learning Skill Curation for Self-Evolving Agents (NeurIPS 2026 · similarity 0.86)