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

WALT: Web Agents that Learn Tools

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 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.

WALT reverse-engineers website-provided functionality (search, filter, post, edit) into validated, callable tools discovered offline by a browser agent and a tool-construction agent, letting web agents act with fewer steps on VisualWebArena and WebArena.

Problem

Web agents relying on step-by-step UI interaction and heavy LLM reasoning are brittle under dynamic layouts and long horizons, and prior skill-discovery methods mine only successful trajectories or hypothesize skills that are brittle UI sequences.

Approach

Instead of agent-induced action sequences, expose website-designed functionality as deterministic high-level tools: a browser agent demonstrates functionality, a tool-construction agent turns traces into action scripts with validated input schemas and URL-parameter promotion, and a test agent validates each tool before registration.

Key results

• WALT reaches 52.9% average success on VisualWebArena (64.1% on Classifieds) and 50.1% on WebArena, versus 40.4% for ASI on WebArena.
• On VisualWebArena Classifieds with GPT-5-mini, discovered tools raise success from 57.5% to 61.5% (text DOM, self-verify) and cut average steps from 8.9 to 6.5.
• Adding multimodal DOM parsing and external verification to discovered tools reaches 64.1% SR at 7.0 steps, versus 57.5% and 8.9 steps for the no-tool baseline.

Surprising findings

• Agentic steps are rare: only 3 of 9 discovered Classifieds tools contain at least one agentic step.
• Self-reported success by web agents suffers from agreement bias, where failed trajectories are often rationalized as successful, motivating external verification.

Limitations

• Offline per-website discovery and validation cost is significant.
• Highly dynamic interfaces, CAPTCHAs, and anti-automation can block URL promotion and reduce determinism.
• Evaluated only on two research benchmarks; external validity to enterprise apps untested.

Experimental setup

Datasets: VisualWebArena WebArena

Benchmarks: VisualWebArena WebArena

Metrics: success rate average number of steps

Baselines: SkillWeaver AWM ASI Hybrid Agent SGV Claude Computer-Use Agent GPT-4+CoT MCTS

Builds on: SkillWeaver Agent Workflow Memory ASI WebJudge browser-use

Closest papers by meaning

• A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.89)
• Go-Browse: Training Web Agents with Structured Exploration (ICLR 2026 · similarity 0.88)
• AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents (ICLR 2025 · similarity 0.87)
• Prune4Web: DOM Tree Pruning Programming for Web Agent (AAAI 2026 · similarity 0.87)
• REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.87)
• WebPilot: A Versatile and Autonomous Multi-Agent System for Web Task Execution with Strategic Exploration (AAAI 2025 · similarity 0.87)