Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR; dynamic outline structures mitigate static-pipeline failures in deep research agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

WebWeaver is a dual-agent framework for open-ended deep research (OEDR) that pairs a planner (iterative search + outline co-evolution) with a writer (section-wise, citation-grounded synthesis from a memory bank). The planner's dynamic research cycle and the writer's hierarchical retrieval address two distinct failure modes of prior work—stale static outlines and long-context attentional collapse—yielding SOTA on DeepResearch Bench, DeepConsult, and DeepResearchGym.

Problem

Existing open-ended deep research agents fail in two ways: (1) static pipeline decoupling—either outline-first or search-first approaches that cannot adapt once started; and (2) monolithic generation—feeding all 100k+ tokens of raw evidence into the writer's context, causing hallucinations, "lost in the middle" degradation, and cross-sectional contextual bleeding. Both problems compound on citation accuracy, where prior SOTA agents lag well behind human reference answers.

Method

WebWeaver uses two ReAct agents sharing a structured memory bank:

Planner — Dynamic Research Cycle. Operates an iterative loop over three actions: search, write_outline, and terminate. After each search, an LLM (1) distills a query-relevant summary fed back into context, and (2) extracts verifiable evidence (quotes, data points) stored in the memory bank under unique IDs. The planner then rewrites/expands the outline, embedding <citation>id_N</citation> tags that anchor each section to specific memory entries. This search↔outline feedback loop repeats until the outline is deemed comprehensive.

Writer — Memory-Grounded Hierarchical Synthesis. Receives the finalized citation-annotated outline and the memory bank. For each section it executes: retrieve (pulls only the IDs cited in that section), a think reasoning step (synthesizes evidence into a narrative), write (produces prose inside <write> tags), then prunes those source materials from context before moving to the next section. This retrieval-and-prune cycle prevents context overflow and cross-sectional interference.

Both agents use the ReAct trajectory formalism: $\(H_T = (\tau_0, a_0, o_0, \ldots, \tau_i, a_i, o_i, \ldots, \tau_T, a_T)\)$ terminating via a <terminate> token.

A high-quality SFT dataset WebWeaver-3k is distilled from the framework to fine-tune smaller models.

Key Contributions

  • Dynamic co-evolution loop coupling search and outline optimization, replacing one-shot or fixed-order paradigms.
  • Citation-anchored memory bank decoupling evidence storage from LLM context, enabling targeted per-section retrieval.
  • Retrieval-and-prune writer that eliminates cross-sectional contextual bleeding and scales to 20k+ token reports.
  • State-of-the-art on three OEDR benchmarks, including outperforming proprietary systems (doubao-research, Claude-research, openai-deepresearch) on DeepResearch Bench.
  • WebWeaver-3k SFT dataset enabling smaller open-weight models to approach expert-level OEDR performance.

Results

  • DeepResearch Bench (RACE overall): WebWeaver (Claude-sonnet-4) = 50.48, WebWeaver (Qwen3-235b) = 50.80; vs. Gemini-2.5-pro-deepresearch = 49.71, openai-deepresearch = 46.45, Claude-research = 45.00.
  • DeepResearch Bench (FACT citation accuracy): WebWeaver (Claude-sonnet-4) = 92.13% effective citation accuracy with 216.99 effective citations; vs. Gemini-2.5-pro-deepresearch at 78.30% / 165.34 citations.
  • DeepConsult win rate: WebWeaver (Claude-sonnet-4) wins 66.16%, ties 12.14%, loses 21.68%; WebWeaver (Qwen3-235b) wins 67.69%, vs. Gemini-2.5-pro at 61.27% win.
  • DeepResearchGym avg. score: WebWeaver (Claude-sonnet-4) = 96.74, WebWeaver (Qwen3-235b) = 95.57; vs. Gemini-2.5-pro = 96.02 (essentially tied at top).
  • WebWeaver (Qwen3-30b, small model) = 50.48 RACE, substantially below larger variants; demonstrates the SFT distillation pathway is viable but not yet competitive at the 30b scale.

Limitations

  • Relies on a separate LLM (GPT-oss-120b) for URL selection, summarization, and evidence extraction within the search tool—adding cost and a non-trivial external dependency.
  • The retrieval-and-prune strategy requires the planner to embed correct citation IDs; errors in citation mapping propagate to the writer's retrieval.
  • Benchmarks are English-centric; generalization to multilingual OEDR is undemonstrated.
  • WebWeaver-3k ablation is deferred to Appendix C (not included in the provided text); distillation effectiveness for small models is asserted but not fully detailed in the main paper.
  • Cost analysis is in Appendix F (not provided); per-query cost may be prohibitive relative to single-pass approaches.

Relevance to Harnesses / Meta-Harnesses

WebWeaver is a textbook example of a multi-agent harness where control flow (plan↔search loop, terminate condition, section-wise writer loop) is encoded structurally rather than left to an unconstrained LLM. The planner↔writer division with a shared memory bank is a harness-level design pattern—analogous to orchestrator/executor splits seen in meta-harness frameworks—where each agent has a bounded action space and explicit state handoff (the citation-annotated outline). The retrieval-and-prune mechanism is a concrete solution to a problem all long-horizon harnesses face: context budget management across sub-tasks. The WebWeaver-3k distillation pipeline further models how a harness can generate structured training data to compress expert-level multi-step reasoning into smaller deployable agents.