SurveyWeave: An Agentic Framework for Automated Survey Generation
Organizations: University of Washington · Meta Reality Labs · University of Washington Tacoma
Abstract
Surveys help researchers keep up with fast-moving fields, but they take weeks to months to write. Automated survey generation (AutoSurvey) can alleviate this situation, but it is challenging to draw on hundreds to thousands of precisely scoped papers, choose among many valid structures, and stay faithful to its sources while offering insights. We present SurveyWeave, an agentic architecture that decomposes the problem into four tasks coordinated by three agents, each reviewing others' work and conducting revisions based on feedback. We also introduce WeaveEval, a five-dimension evaluation grounded in prior surveys. A comprehensive comparison shows that SurveyWeave outperforms Codex by 8 points and the strongest AutoSurvey baseline by 6 points.
Figures & tables
| Collection | Outline | Cate. | Surv. | Architecture | |||||||
| System | LLM filtering | Bkg papers | Taxo-nomy based | Exter. evid. | Multi outline | Quality based | Fig/ Table gen | Agentic | Ent. based agents | Cross agent rev. | Back revision |
| Pipelined approaches | |||||||||||
| AutoSurvey Wang et al. (2024) | ✗ | ✗ | ✗ | ✗ | ◐ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ |
| SurveyX Liang et al. (2025) | ✓ | ✗ | ◐ | ◐ | ✗ | ✗ | ✓ | ✗ | ✗ | ✗ | ◐ |
| SurveyForge Yan et al. (2025) | ✗ | ✗ | ✗ | ◐ | ◐ | ✓ | ✗ | ◐ | ✗ | ✗ | ✗ |
| LLM MapReduce-V2 Wang et al. (2025) | ✗ | ✗ | ◐ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ◐ | ✗ |
| Evidence | Coverage | Synthesis | New insight | Presentation | Overall | |||||||||||||||||
| System | E1 | E2 | E3 | E | C1 | C2 | C | S1 | S2 | S3 | S | N1 | N2 | N | P1 | P2 | P3 | P4 | P | Q-txt | Q | DAS-Eval |
| Released surveys † | ||||||||||||||||||||||
| DAS | 63.5 | 89.6 | 93.3 | 77.5 | 77.8 | 59.1 | 68.5 | 70.7 | 52.7 | 69.6 | 64.1 | 46.4 | 83.9 | 55.8 | 75.0 | 51.8 | 46.4 | 58.9 | 58.0 | 65.6 | 64.8 | 4.45 |
| InteractiveSurv. | 31.4 | 17.0 | 62.1 | 35.5 | 81.8 | 38.5 | 60.1 | 63.7 | 52.7 | 75.9 | 62.9 | 36.6 | 91.1 | 50.2 | 73.2 | 56.2 | 25.9 | 31.2 | 46.7 | 52.5 | 51.1 | 3.62 |
| General agent | ||||||||||||||||||||||
| Codex | 85.8 | 64.4 | 96.5 | 83.1 | 91.4 | 87.5 | 89.5 | 76.7 | 84.7 | 99.3 | 85.2 | 68.8 | 99.3 | 76.4 | 76.4 | 63.9 | 54.2 | 66.7 | 65.3 | 80.6 | 79.9 | 4.51 |
| Variant | Evi | Cov | Syn | New | Pres | Q |
|---|---|---|---|---|---|---|
| SurveyWeave | 97.8 | 91.3 | 91.0 | 81.3 | 81.3 | 88.5 |
| One-way pipeline | 91.0 | 88.4 | 88.1 | 77.5 | 75.0 | 84.0 |
| One-shot outline | 90.9 | 86.7 | 82.5 | 80.0 | 72.5 | 82.5 |
| Collection-only Paper | 90.1 | 84.0 | 83.6 | 77.5 | 73.8 | 81.8 |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Hyperparameter | Value |
|---|---|
| Paper collection | |
| Semantic / keyword queries | 3–4 / 4–8 |
| Semantic hits per query | 1,500 |
| Classifier round 1 (top / random) | 200 / 200 |
| Classifier round 2 (outside queries) | 100 |
| Background budget coefficient | 16 |
| System | Code | Papers available | Full text | Figures / tables | Length setting |
|---|---|---|---|---|---|
| SurveyWeave | Ours | Our corpus, collected per topic, and the Web | Yes | Generated figures; tables | 15,000–22,000 words (instruction) |
| TargetedRAG | Ours, following DAS-Bench | 300 papers from SurveyWeave ’s paper collection | No | TikZ figures; tables | 15,000–22,000 words (prompt) |
| AutoSurvey | Released, commit 5e8f389 | Titles and abstracts of our corpus | No | Text only | 7 sections, 3–4 subsections each, about 500 words per subsection |
| SurveyForge | Released, commit 9114a0b | Titles and abstracts of our corpus; released index of 12,756 human-written surveys | No | Text only | 7 body sections; per subsection, words (at least 500) |
| SurveyAgent-HKA | Released, commit 4f544b4 | Abstracts of our corpus; live arXiv search | No | Text; no figure step | at most 50 subsections, 300–500 words each |
| Codex | Codex agent, our task prompt | Open Web | Yes | Its own tools; no image generation | 15,000–22,000 words (prompt) |
| ID | Topic | Benchmark | Subfield |
| DAS-009+cs_03 | Diffusion and Flow-Based Models for Controllable Image Generation | Both | Vision |
| DAS-012+cs_07 | Continual Learning and Model Editing for Foundation Models | Both | ML methods |
| DAS-001 | Tool Learning and Function Calling for LLM Agents | DAS-Bench CS | NLP |
| DAS-002 | Memory and Long-Context Mechanisms for Long-Horizon LLM Agents | DAS-Bench CS | NLP |
| DAS-003 | Retrieval-Augmented Generation for Large Language Models | DAS-Bench CS | IR and RecSys |
| DAS-004 | Planning and Self-Reflection in Large Language Model Reasoning | DAS-Bench CS | NLP |
| Cost (USD) | Time (min) | |||
|---|---|---|---|---|
| System | Mean | Range | Mean | Range |
| TargetedRAG | 0.40 | 0.19–0.59 | 11.8 | 7–16 |
| Codex | 1.05 | 0.35–2.01 | 25.7 | 6–56 |
| AutoSurvey | 4.39 | 3.13–6.07 | 15.5 | 8–34 |
| SurveyForge | 7.60 | 5.11–10.55 | 12.8 | 7–25 |
| SurveyAgent-HKA | 7.72 | 4.59–10.81 | 78.9 | 26–271 |
| Agent | Own calls | With sub-agents |
|---|---|---|
| Survey | 6.00 | 12.17 |
| Paper | 6.43 | 11.73 |
| Outline | 0.73 | 1.20 |
| Host | 0.78 | 0.78 |
| Figure generation | – | 1.85 |
| Paper summaries | – | 0.10 |
| ID | Name | What it measures | Scoring | Weight |
| E1 | Factual accuracy | Sampled factual claims are fully supported by the full text of the cited papers | Claim check | 0.50 |
| E2 | Attribution completeness | Statements that need a citation have one that points to the intended work | Claim check | 0.25 |
| E3 | Citation relevance | Each attached citation is relevant to the statement it supports | Claim check | 0.25 |
| C1 | Core explanation | The topic’s core concepts and mechanisms are explained | Checklist | 0.50 |
| C2 | Evaluation methodology | Metrics, benchmarks, protocols, and how to read results are explained | Checklist | 0.50 |
| S1 | Relationship coverage | Expected cross-study relationships and trade-offs are stated | Checklist | 0.40 |
| WeaveEval Q | DAS-Eval | |||||
|---|---|---|---|---|---|---|
| vs. | 95% CI | W/T/L | 95% CI | W/T/L | ||
| SurveyForge | +6.3 | [5.2, 7.3] | 33/0/3 | +0.53 | [0.49, 0.57] | 36/0/0 |
| TargetedRAG | +6.4 | [5.4, 7.6] | 36/0/0 | +0.07 | [ 0.00, 0.15] | 15/7/14 |
| Codex | +7.7 | [5.7, 9.5] | 31/0/5 | +0.07 | [0.02, 0.12] | 22/6/8 |
| AutoSurvey | +9.8 | [8.4, 11.2] | 36/0/0 | +0.48 | [0.43, 0.53] | 36/0/0 |
| SurveyAgent-HKA | +13.1 | [11.7, 14.4] | 36/0/0 | +0.54 | [0.49, 0.60] | 36/0/0 |
| Score | Label | Description |
|---|---|---|
| 5 | Excellent | I’d point someone new to this topic straight to it. Accurate, well organized, with insight beyond summarizing papers. |
| 4 | Good | Accurate and well organized, but mostly summarizes papers without much insight. |
| 3 | Adequate | Useful and largely accurate, with minor gaps or weaker sections. |
| 2 | Weak | Worth reading, but with noticeable gaps, shallow parts, or organization problems. |
| 1 | Not useful | Limited value: major gaps, errors, hard to follow, and may misguide readers. |
Autonomous Research Agents
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.
See all 346 papers