DeepWeaver - Bridging the Evidence Synthesis Gap in Open-Ended Question Answering

Sep 15, 2026

Uploaded by Asher

Retrieve-then-generate pipelines can answer open-ended questions, yet retrieval alone fails to ensure comprehensive, well-cited responses because evidence must be reorganized into coherent synthesis. This document identifies an evidence synthesis gap: direct generation underuses evidence, misaligns citations, and compresses diverse information into shallow summaries. It proposes DeepWeaver, which weaves noisy retrieved evidence into structured Thought Block Chains, supporting iterative revisions and new claim discovery, and evaluates it on LoQA and DeepResearch Bench.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages40

Reading Time1 h 41 min

AccessFree to download

Introduction

Evidence synthesis gap in retrieval-then-generate pipelines

DeepWeaver: Thought Block Chains and evidence weaving

Evaluation on LoQA and DeepResearch Bench

DeepWeaver: Bridging the Evidence Synthesis Gap in Open-Ended Question Answering Xujia Wang Yizhe Zhang Bin Xu† Lei Hou Juanzi Li Department of Computer Science and Technology, Tsinghua University [wang-xj22@mails.tsinghua.edu.cn](wang-xj22@mails.tsinghua.edu.cn) arXiv : 2608 . 18988v2 [cs .CL] 6 Sep 2026 Abstract Retrieve-then-generate pipelines are commonly used to produce deep-research answers for open-ended questions, but retrieval alone is insufficient: LLMs must organize noisy and fragmented evidence into comprehensive, wellcited answers. We refer to this process as evidence synthesis. However, direct generation often underuses evidence, misaligns citations, and collapses diverse information into shallow summaries, exposing an evidence synthesis gap between retrieval and generation. Thus, we propose DeepWeaver, a novel framework that weaves noisy retrieved evidence into comprehensive answers by maintaining Thought Block Chains (TBCs), a structured representation that groups claims, salient information, keywords, and supporting evidence. DeepWeaver uses subordinate TBCs to inspect residual evidence, commit TBC revisions, and discover new claims before final generation. We evaluate DeepWeaver on open-ended QA over both knowledge bases and the web, and introduce LoQA, a high-density benchmark for evidence synthesis. Across LLMs, DeepWeaver improves content sufficiency, citation grounding, and detail preservation on LoQA, while achieving deeper insights and higher citation quality on DeepResearch Bench. These results show that evidence weaving is effective for bridging retrieval and generation in open-ended QA. Our code is available at this URL. 1 Introduction Large language models (LLMs) are increasingly being used to generate research-style answers grounded in external evidence (Menick et al., 2022 ; Asai et al., 2024a) . For domain-specific openended questions and deep-research-style tasks, answering a query may require integrating evidence from hundreds of retrieved passages, web pages, books, or technical documents. As a result, evidence retrieval and synthesis are central to both † Corresponding author. Figure 1: DeepWeaver bridges the evidence synthesis gap between retrieval and generation by weaving noisy evidence into structured Thought Block Chains (TBCs) . Unlike naïve retrieval-augmented generation, DeepWeaver synthesizes noisy evidence fragments into answers with well-supported woven claims, fine-grained details, and grounded citations. retrieval-augmented generation (RAG) pipelines (Lewis et al., 2021 ; Izacard and Grave, 2021) and web-based deep-research agents (Nakano et al., 2022 ; Yao et al., 2023) . These systems share a common workflow: retrieve evidence, place it into the context, and prompt the model to synthesize a comprehensive answer (Huang and Huang, 2026) . Although prior work has extensively studied how to improve evidence retrieval quality (Han et al., 2025 ; Guo et al., 2025b ; Sun et al., 2025), retrieval itself is merely the first step: enabling LLMs to effectively use extensive, noisy, and fragmented evidence to generate comprehensive answers remains a key challenge. To better integrate upstream evidence, context-refinement methods (Xu et al., 2024a ; Jin et al., 2025b) compress, filter, or restructure passages before generation, while divide-and-conquer methods (Zhang et al., 2024b ; Zhou et al., 2025) split long contexts into segments and aggregate intermediate outputs. End-to-end deep-research systems (Wang et al., 2024c ; Tao et al., 2025 ; Li et al., 2026) go further by automating evidence organization and long-form answer writing with agent memory and outline-guided generation. However, prior methods overestimate LLMs’ability to fully exploit input evidence to derive indepth, diverse claims. By primarily optimizing evidence presentation and outline-guided writing (Xu and Peng, 2025), these approaches often fall short in deeply synthesizing retrieved evidence in

FAQs

What benchmarks and evidence sources are used to evaluate DeepWeaver?

DeepWeaver is evaluated on LoQA, which uses evidence from a Chinese expert knowledge base with many books and large text volume, and on web-based deep-research questions from DeepResearch Bench.

Sep 15, 2026

Uploaded by Asher