Year: 2026
Type: conference-paper
Abstract: Multimodal BrowseComp tasks require agents to combine perception, tool use, and longhorizon reasoning over dynamic web content, challenging their ability to handle compositional structure, open-world uncertainty, and multimodal integration across extended interactions.Crucially, real-world multimodal browsing involves three distinct information-flow patterns: text-only, image-to-text, and textto-image, yet existing data construction methods cover only the text-only and image-to-text patterns, leaving text-to-image largely unaddressed and limiting agent generality and robustness.We introduce UNIBROWSE, a unified data pipeline that for the first time simultaneously generates training data covering all three patterns, augments curated knowledge graphs with live web retrieval for improved fidelity, and introduces a novel metric of exploration degree to filter low-signal instances for efficient reinforcement learning.Through this pipeline, we produce high-quality cold-start tool-use trajectories and explorationrich QA pairs, and train a 35B-scale agent via supervised fine-tuning and explorationaware RL.The resulting UNIBROWSE agent achieves state-of-the-art performance on multimodal BrowseComp benchmarks, attaining an average accuracy of 54.4 across five diverse benchmarks-an improvement of 10.5 points over its base model Qwen3.5-35B-A3Bandsurpassing serveral closed-source agent workflows such as GPT-5 (42.9),Gemini-2.5Pro (44.8), and Gemini-2.5Flash (41.3).We will release our code and dataset at https: //github.com/lemon-prog123/UniBrowse. UniBrowse-35B-A3BGemini-2.5 Pro GPT-5 Kimi K2.5 Qwen3.
Institution: Peking University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
Awards: G3336050986, G4128466783
OpenAlex ID: W7168248013