Abstract

Recent advances in deep-research systems have demonstrated the potential forAI agents to autonomously discover and synthesize knowledge from externalsources. In this paper, we introduce WebResearcher, a novel framework forbuilding such agents through two key components: (1) WebResearcher, aniterative deep-research paradigm that reformulates deep research as a MarkovDecision Process, where agents periodically consolidate findings into evolvingreports while maintaining focused workspaces, overcoming the contextsuffocation and noise contamination that plague existing mono-contextualapproaches; and (2) WebFrontier, a scalable data synthesis engine thatgenerates high-quality training data through tool-augmented complexityescalation, enabling systematic creation of research tasks that bridge the gapbetween passive knowledge recall and active knowledge construction. Notably, wefind that the training data from our paradigm significantly enhances tool-usecapabilities even for traditional mono-contextual methods. Furthermore, ourparadigm naturally scales through parallel thinking, enabling concurrentmulti-agent exploration for more comprehensive conclusions. Extensiveexperiments across 6 challenging benchmarks demonstrate that WebResearcherachieves state-of-the-art performance, even surpassing frontier proprietarysystems.