Alibaba-NLP/WebWatcher-7B

WebWatcher-7B: A Multimodal Deep Research Agent

WebWatcher-7B, developed by Alibaba-NLP, is a 7 billion parameter multimodal agent engineered for deep research, integrating advanced visual-language reasoning with multi-tool interaction. This model introduces a unified framework to tackle complex information gathering and analysis tasks.

Key Capabilities

• Enhanced Visual-Language Reasoning: Combines visual perception with advanced information-gathering abilities for in-depth multimodal reasoning.
• Multi-Tool Interaction: Equipped with tools such as Web Image Search, Web Text Search, Webpage Visit, Code Interpreter, and an internal OCR tool.
• Automated Trajectory Generation: Utilizes an automated pipeline to create high-quality, multi-step reasoning trajectories for efficient training and optimization.
• Superior Performance: Significantly outperforms proprietary baselines and other open-source agents across challenging VQA benchmarks.

Performance Highlights

• Complex Reasoning (HLE-VL): Achieved a Pass@1 score of 13.6%, surpassing GPT-4o (9.8%) and Gemini2.5-flash (9.2%).
• Information Retrieval (MMSearch): Demonstrated exceptional retrieval accuracy with a Pass@1 score of 55.3%, outperforming Gemini2.5-flash (43.9%) and GPT-4o (24.1%).
• Knowledge-Retrieval Integration (LiveVQA): Scored 58.7% Pass@1, exceeding Gemini2.5-flash (41.3%) and GPT-4o (34.0%).
• Information Optimization and Aggregation (BrowseComp-VL): Dominated with an average score of 27.0%, more than doubling the performance of models like GPT-4o (13.4%) and Gemini2.5-flash (13.0%).

Good for

• Advanced visual search and information aggregation.
• Complex, multi-step reasoning tasks requiring both visual and textual understanding.
• Applications demanding robust tool-use capabilities for web-based research.
• Benchmarking against challenging multimodal datasets like HLE-VL and BrowseComp-VL.