Research map / Multi-agent systems and games / LLM agents
Safe and Scalable Web Agent Learning via Recreated Websites
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 VeriEnv, which uses a coding agent (GPT-5.2) to clone 149 real websites into executable synthetic environments with SDK-based verifiable tasks and judges, then trains web agents (Qwen3-4B, LLaMA-3.2-3B) via rejection fine-tuning on verified trajectories.
Problem
Training web agents on real websites is unsafe, hard to reset, and yields unverifiable tasks and rewards that rely on LLM judges, which makes self-evolving learning unstable.
Approach
Treat language models as environment creators: a coding agent clones websites (code, database, and a Python SDK for internal state access), and tasks come with executable validation programs that give deterministic binary rewards for rejection fine-tuning.
Key results
- On WebArena-Lite, VeriEnv improves Qwen3-4B from 7.88 to 13.94 (+6.06) and LLaMA-3.2-3B-Instruct from 3.03 to 12.73 (+9.70) total success rate.
- On Mind2Web-Online, VeriEnv raises Qwen3-4B from 13.18 to 20.45 (+7.27) and LLaMA-3.2-3B-Instruct from 11.36 to 24.55 (+13.19).
- Human evaluation of 149 cloned sites reports 90% functional correctness, 4.7/5 visual rating, 90% task executability, and 76% judge correctness.
Surprising findings
- ADP improves LLaMA-3.2-3B-Instruct but degrades Qwen-based models, which the authors attribute to heterogeneous dataset mixing and action-format mismatch.
- On Mind2Web-Online, VeriEnv's gain for LLaMA-3.2-3B (+13.19) is slightly below ADP's (+12.73) on the Hard split VeriEnv does not beat ADP's easy split, showing mixed per-difficulty results.
Limitations
- Judge correctness is only 76%, and database reset issues (unpreserved random seeds) cause some verifier errors.
- 39 of 136 websites failed to be implemented, mainly due to incomplete setup and infrastructure issues such as port conflicts and CORS.
- Cloned sites are not identical to originals, and multimedia-heavy sites (arXiv, YouTube) are hard to reproduce.
Experimental setup
Datasets: Mind2Web WebArena Mind2Web-Online
Benchmarks: WebArena-Lite Mind2Web-Online
Metrics: success rate
Baselines: GPT-4o-mini GPT-4o Claude-3.5-Sonnet Synatra Agent Data Protocol PAE Browser-Use-GPT-4o
Builds on: WebRL Synatra Agent Data Protocol PAE WebArena
Closest papers by meaning
- WebArena: A Realistic Web Environment for Building Autonomous Agents (ICLR 2024 · similarity 0.88)
- REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.88)
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.88)
- WebGym: Scaling Training Environments for Long-Horizon Visual Web Agents with Realistic Tasks (CVPR 2026 · similarity 0.88)
- Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration (CVPR 2026 · similarity 0.87)
- A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.