Researchers have proposed a new framework called DeepWeaver to address the issue of evidence synthesis in open-ended question answering. The current approach of using retrieval-then-generation pipelines often falls short, as large language models (LLMs) struggle to organize and present noisy evidence into comprehensive answers. To bridge this gap, DeepWeaver uses a structured representation called Thought Block Chains (TBCs), which groups claims, information, keywords, and supporting evidence. This framework allows for the inspection of residual evidence, revisions of TBCs, and discovery of new claims before final generation. The authors evaluate DeepWeaver on open-ended QA tasks over knowledge bases and the web, introducing a new benchmark called LoQA. --- Why it matters: This matters to researchers in AI because it addresses a significant challenge in open-ended question answering: how to effectively synthesize evidence from multiple sources into coherent answers. DeepWeaver's approach has the potential to improve the accuracy and reliability of LLMs in this task, which is crucial for applications such as research summarization and decision-making. Source: https://arxiv.org/abs/2608.18988