๐ LLM
AI generated
WebGym: Open-Source Environment for Training Visual Web Agents
## WebGym: A New Environment for Training Visual Web Agents
WebGym has been introduced, an open-source environment designed for training realistic visual web agents. This tool stands out for its vast range of tasks, nearly 300,000, covering real-world websites and diverse difficulty levels.
## Scalability and Performance
To scale reinforcement learning (RL), WebGym introduces a high-throughput asynchronous rollout system, optimized for web agents. This system speeds up trajectory sampling by 4-5x compared to standard implementations. Training a vision-language model, Qwen-3-VL-8B-Instruct, on WebGym led to an increase in the success rate on an out-of-distribution test set from 26.2% to 42.9%, surpassing agents based on proprietary models such as GPT-4o and GPT-5-Thinking.
## Implications
WebGym's ability to improve performance on websites never seen during training represents a significant step forward in the development of robust and adaptable visual web agents. This is particularly important because many previous works focused on tasks on websites already seen during training.
๐ฌ Commenti (0)
๐ Accedi o registrati per commentare gli articoli.
Nessun commento ancora. Sii il primo a commentare!