RAG／深度研究代理
DeepWeaver Weaves Evidence Before Generating Answers, Raising DeepSeek-V4-Flash’s Argument Sufficiency from 75.4% to 86.5%
DeepWeaver uses a Thought Block Chain to organize claims, key information, and sources, then repeatedly checks for evidence that has not yet been used. It improves long-form responses and citation metrics across several models, but the new benchmark focuses on Chinese-language water environment questions, and the code has yet to be released.
A team at Tsinghua University has introduced DeepWeaver, an attempt to bridge the gap between “finding information” and “writing reliable answers” in deep research systems. Conventional RAG systems feed large volumes of retrieved passages directly into a long context, but experiments show that adding more noisy evidence does not necessarily outperform sampling only 120 passages. Models may overlook details, merge incompatible claims, or attach citations to the wrong sentences.
Before generating the final answer, DeepWeaver maintains a Thought Block Chain (TBC). Each block links a claim with summarized information, keywords, and supporting evidence. The system first creates a draft, then uses a subordinate TBC to scan passages that have not yet been covered. Finally, a commit step merges new claims and discards duplicate content. Two refinement rounds produced the best results, while a third yielded no clear additional gains.
On the newly created LoQA benchmark, adding DeepWeaver to DeepSeek-V4-Flash increased evidence recall from 24.2 to 31.6, argument sufficiency from 75.4% to 86.5%, and the number of relevant citations from 31.0 to 44.6. Qwen3.5-122B-A10B and DeepSeek-V3.2 also showed consistent improvements. The web research evaluation continued to use WebWeaver’s search engine unchanged to isolate the effects of evidence synthesis. DeepWeaver achieved an effective citation score of 60.13, compared with 26.74 for WebWeaver.
These gains do not come entirely without cost. When running Qwen3-30B-A3B on two A100 GPUs, DeepWeaver processed an average of 3.16 million input tokens, more than WebWeaver’s 2.35 million. However, intermediate output fell from 307,000 to 94,000 tokens, local runtime decreased from 18.5 minutes to 14.2 minutes, and estimated costs were similar at $0.313 and $0.303, respectively. Engineering teams should note that LoQA covers only the Chinese-language water environment domain, several quality scores were generated by LLM judges, and the GitHub repository currently contains only documentation stating that the code will be updated later. The complete workflow therefore cannot yet be reproduced independently.