Research map / Multi-agent systems and games / LLM agents

WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces WebWeaver, a dual-agent (planner + writer) framework that interleaves web search with outline optimization and writes reports section by section via citation-grounded memory retrieval, reaching SOTA on DeepResearch Bench, DeepConsult and DeepResearchGym.

Problem

Open-ended deep research requires synthesizing web-scale evidence into accurate, well-cited long reports, but static plan-then-search pipelines and monolithic long-context generation cause outdated outlines, redundant context, hallucinations and poor citation accuracy.

Approach

A planner that co-evolves the outline and search in a dynamic loop, binding outline sections to evidence IDs in a memory bank, and a writer that retrieves only the evidence cited by each section and prunes it after writing, instead of feeding all evidence into one context.

Key results

• WebWeaver (Claude-sonnet-4) reaches 50.58 RACE overall and 93.37% citation accuracy on DeepResearch Bench, versus 49.71 and 78.30% for Gemini-2.5-pro-deepresearch.
• WebWeaver (Claude-sonnet-4) achieves a 66.86% win rate on DeepConsult against openai-deepsearch and 96.77 average score on DeepResearchGym.
• SFT of Qwen3-30B-A3B on WebWeaver-3k raises DeepResearchGym score from 77.27 to 90.89 and citation accuracy from about 25% to 85.90%.

Surprising findings

• Hierarchical writing outperforms brute-force LongWriter-style writing on insight and readability, not only on citation accuracy.
• The untuned Qwen3-30B-A3B base has very low citation accuracy (25%), which SFT raised sharply.

Limitations

• Main results depend on proprietary API models (Claude-sonnet-4, GPT-4.1 judges), limiting reproducibility and cost accessibility.
• LLM-as-judge evaluation with Gemini or GPT models may bias scores; the outline ablation uses GPT-4.1-mini judge only.

Experimental setup

Datasets: DeepResearch Bench DeepConsult DeepResearchGym Researchy Questions WebWeaver-3k

Benchmarks: DeepResearch Bench DeepConsult DeepResearchGym

Metrics: RACE overall score comprehensiveness insight instruction-following readability effective citations citation accuracy win/tie/lose rate clarity/depth/balance/breadth/support/insightfulness average

Baselines: WebShaper (32B) langchain-open-deep-research doubao-research kimi-research Claude-research openai-deepresearch Gemini-2.5-pro-deepresearch LongWriter-style brute-force writing

Builds on: ReAct LongWriter WebShaper

Closest papers by meaning

• From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (NeurIPS 2026 · similarity 0.89)
• WebThinker: Empowering Large Reasoning Models with Deep Research Capability (NeurIPS 2025 · similarity 0.88)
• DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence (ICLR 2026 · similarity 0.87)
• SciResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery (NeurIPS 2026 · similarity 0.86)
• MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (NeurIPS 2026 · similarity 0.86)
• WebWatcher: Breaking New Frontiers of Vision-Language Deep Research Agent (ICLR 2026 · similarity 0.86)