AI for Auto-Research: A Survey

Awesome AI Auto-Research Team A comprehensive practitioner's guide for using AI-assisted tools across the complete academic research lifecycle.

arXiv PDF GitHub HuggingFace Paper
AI Auto-Research Teaser
Crossing a threshold

AI-assisted research is crossing a threshold. Fully automated systems can now generate research papers for as little as $15, while long-horizon agents execute experiments, draft manuscripts, and simulate critique with minimal human input — yet under scientific pressure, even frontier LLMs still fabricate experiment results, miss hidden errors, and fail to judge novelty reliably.

AI auto-research across the complete lifecycle: four phases and eight stages.
AI auto-research across the complete lifecycle. Four phases and eight stages: Creation (ideation, literature, code & experiments, tables & figures), Writing (paper writing), Validation (peer review, rebuttal & revision), and Dissemination (posters, slides, videos, social media, project pages, and interactive paper agents).

Across this lifecycle we identify a sharp, stage-dependent boundary between reliable assistance and unreliable autonomy. The core challenge is no longer whether AI can produce the forms of research, but whether it can preserve the substance — evidence, judgment, provenance, and accountability. We argue that human-governed collaboration is the most credible deployment paradigm, and that effective systems converge on layered architectures combining exploration, execution, and verification.

250+
Papers Surveyed
8
Research Stages
4
Lifecycle Phases
5
Central Findings


The Academic Research Lifecycle

Four Phases, Eight Stages

We organize the academic research lifecycle as eight interconnected stages grouped into four epistemological phases. Each phase serves a distinct function in producing, scrutinizing, and communicating scientific knowledge.

Phase 1: Creation
Phase 1 · Stages 1–4

Creation

The stages through which a research contribution is materially produced.

From spark to substance: ideation grounded in the literature, executed in code, and rendered as figures and tables. This is where AI's promise — and its capability boundary — is most visible.

S1: Idea Generation
Stage 1

Idea Generation

Know more

Generating, refining, and evaluating research hypotheses. Systems span direct LLM prompting, retrieval-augmented and knowledge-graph generation, multi-agent collaboration, and learned quality signals. The central challenge: LLMs can produce ideas that appear novel and well-motivated, yet often struggle to generate ones that remain feasible, distinctive, and impactful after execution.

Sub-topics
LLM-based generation KG-driven Trend-driven Multi-agent ideation Novelty assessment Human-AI co-ideation
Representative methods
AI Scientist ResearchAgent SciMON VirSci SciAgents IdeaSynth MOOSE-Chem
S2: Literature Review
Stage 2

Literature Review

Know more

Retrieving, synthesizing, and organizing prior work into coherent research contexts. Compared with idea generation, this stage is more grounded and externally verifiable, making it one of the fastest-maturing areas in AI-assisted research. Yet faithful citation, coverage completeness, and multi-paper relational reasoning remain difficult.

Sub-topics
Semantic retrieval Survey generation Deep Research Citation graph Hierarchical org. Related work gen.
Representative methods
PaperQA2 AutoSurvey STORM SurveyX SurveyForge OpenScholar
S3: Coding & Experiments
Stage 3

Coding & Experiments

Know more

Translating ideas into executable code, running experiments, and analyzing empirical results. The challenge is not whether LLMs can write plausible code, but whether they can produce semantically correct research implementations, execute meaningful experiments, and interpret results reliably — performance still drops sharply on genuinely novel research code.

Sub-topics
Paper-to-Code Experiment design Auto execution Result analysis Lab automation Benchmarking
Representative methods
PaperCoder MLAgentBench AIDE R&D-Agent ChemCrow Coscientist
S4: Tables & Figures
Stage 4

Tables & Figures

Know more

Constructing method diagrams, result plots, comparison tables, mathematical formulas, and algorithmic illustrations. Despite their importance in daily research practice, this stage remains comparatively underdeveloped — current tools serve as assistants rather than autonomous producers, and AI-generated figures frequently require human modification for domain-specific symbols and paper-specific visual languages.

Sub-topics
Method diagrams Result plots Table generation LaTeX / TikZ SVG generation Visual feedback
Representative methods
AutoFigure MatPlotAgent PlotGen ChartGPT AutomaTikZ DeTikZify
Phase 2: Writing
Phase 2 · Stage 5

Writing

Organizing the outputs of Creation into a formal scholarly manuscript for communication and external scrutiny.

Writing is not a formatting step — it is a rhetorical and evidential organization process that requires distinct AI capabilities from those used to produce code, experiments, or figures.

S5: Paper Writing
Stage 5

Paper Writing

Know more

Drafting, editing, polishing, and structuring academic manuscripts. AI assistance ranges from grammar correction and citation support to section-level drafting and full-paper generation. The central failure mode is no longer ungrammatical prose but the gap between fluency and argumentative depth; end-to-end autonomous systems have not yet consistently reached major-venue acceptance standards.

Sub-topics
Semi-auto assist Full-auto generation Citation insertion AI text detection LaTeX generation Style & polishing
Representative methods
CycleResearcher ScholarCopilot CoAuthor TeXpert FutureGen FARS
Phase 3: Validation
Phase 3 · Stages 6–7

Validation

External scrutiny — how the community challenges and refines a manuscript.

Together, peer review and rebuttal form the community-facing mechanism through which scientific claims are challenged, defended, and revised — not a bureaucratic afterthought, but the verification layer of research.

S6: Peer Review
Stage 6

Peer Review

Know more

Generating structured reviews, matching reviewers to manuscripts, assessing review quality, and supporting meta-review decisions. These systems aim to assist, not replace, the community's evaluative process — standalone AI review remains unsafe, with persistent leniency bias and vulnerability to adversarial manipulation.

Sub-topics
Review generation Meta-review Reviewer matching Quality assessment Bias detection Review feedback
Representative methods
DeepReviewer MARG REMOR AgentReview OpenReviewer
S7: Rebuttal & Revision
Stage 7

Rebuttal & Revision

Know more

Analyzing reviewer comments, identifying required evidence, drafting evidence-grounded responses, and supporting manuscript revision. This stage connects external critique with additional analysis, clarification, and experimental follow-up — under-served relative to its practical importance, despite being where authors directly negotiate reviewer concerns before final decisions.

Sub-topics
Comment analysis Response generation Evidence retrieval Strategy planning Effectiveness analysis
Representative methods
RebuttalAgent Paper2Rebuttal ReviewerToo ReviewMT Re² Author-in-Loop
Phase 4: Dissemination
Phase 4 · Stage 8

Dissemination

Translating the manuscript into formats accessible to broader audiences.

Posters, slides, videos, project pages, social-media summaries, and interactive agents are independent knowledge artifacts with their own fidelity and trust requirements — not appendages to the paper.

S8: Paper2X
Stage 8

Paper2X

Know more

Converting papers into posters, slides, videos, social media, project pages, and interactive agents. The core bottleneck is trust: researchers may use AI to draft public-facing materials, but remain reluctant to delegate final communication to systems that may distort results, overstate claims, or omit important limitations.

Sub-topics
Paper→Poster Paper→Slides Paper→Video Paper→Social Paper→Web Paper→Agent
Representative methods
Paper2Poster PPTAgent DeepPresenter PresentAgent Paper2Video Paper2Web Paper2Agent
S8.1: Paper2Video
Stage 8.1

Paper2Video

Know more

Generating explanatory videos from papers by synchronizing script, narration, animations, and subtitles across multiple output channels. Video is the most demanding Paper2X format — it must coordinate visual, auditory, and temporal channels simultaneously while remaining faithful to the paper's claims.

Sub-topics
Script generation Narration synthesis Slide animation Subtitle alignment Talking-head video Visual storytelling
Representative methods
Paper2Video Preacher AutoPresent ScholarCast
S8.2: Paper2Slides
Stage 8.2

Paper2Slides

Know more

Automatically distilling paper content into structured presentation decks. Systems must extract key contributions, arrange them into a logical slide flow, and generate visual representations — bridging the gap between dense manuscript prose and audience-ready slides.

Sub-topics
Content extraction Slide structuring Layout generation Figure repurposing Key message distillation Design automation
Representative methods
PPTAgent DeepPresenter PresentAgent Paper2Slides SlideCrafter D2S
S8.3: Paper2Poster
Stage 8.3

Paper2Poster

Know more

Condensing papers into visually compelling research posters, requiring simultaneous summarization and visual design. The core challenge is balancing information density, spatial layout, and aesthetic coherence within a single high-resolution canvas.

Sub-topics
Layout planning Content prioritization Figure adaptation Typography Visual hierarchy Template generation
Representative methods
Paper2Poster PosterLLaVA AutoPoster PosterAgent
S8.4: Paper2Social
Stage 8.4

Paper2Social

Know more

Crafting platform-adapted social media posts from academic papers. Each platform (X, LinkedIn, Reddit, Mastodon) has distinct tone, length, and audience expectations — requiring format-aware generation that simplifies results without distorting them.

Sub-topics
Audience adaptation Tone calibration Thread generation Platform formatting Claim simplification Hashtag generation
Representative methods
Paper2Social TweetPaper PaperDigest SciComm
S8.5: Paper2Agent
Stage 8.5

Paper2Agent

Know more

Building interactive agents that allow readers to query a paper's methodology, results, and implications through conversation. These systems ground responses in specific paper content, enabling exploration beyond the static manuscript.

Sub-topics
Paper grounding Conversational QA MCP integration Citation tracing Demo interfaces Fidelity evaluation
Representative methods
Paper2Agent PaperCopilot DocAgent ScholarBot


Paper Collection

Per-Stage Inventory of Surveyed Methods

A curated, expandable inventory of method discussed in the survey, organized by stage. Click a stage to expand its full table; use the search box to filter by method name, venue, or category.

S1 Idea Generation 40 methods Expand
# Method Venue Category Evaluation Links
1 PosterMELD arXiv'26 Paper2Poster Controllable design diversity with editable poster output
2 SeaSlides arXiv'26 Paper2Slides Semantic abstraction layer for agentic slide generation
LLM Internal Knowledge-Based Generation
3 Chain of Ideas ACL'24 LLM Internal Comparable to human quality; ~$0.50/idea minimum cost.
4 ResearchAgent NAACL'25 LLM Internal Human + model evaluation; academic graph feedback for refinement.
5 SciMON ACL'24 LLM Internal Mitigates shallow novelty via iterative refinement.
6 Idea Gen Agent arXiv'24 LLM Internal 100+ NLP researchers; LLM ideas higher novelty (p<0.05).
7 IRIS ACL'25 LLM Internal MCTS adaptive reasoning; human-in-the-loop platform.
8 Spark ICCC'25 LLM Internal Judge model trained on 600K OpenReview reviews.
9 Rubric Rewards arXiv'25 LLM Internal 70% expert preference; RL with rubric self-grading.
10 DeepInnovator arXiv'26 LLM Internal 80–94% win rates vs. frontier models; 14B parameters.
11 Tree-of-Ideas arXiv'26 LLM Internal Cross-trajectory reasoning over scholarly evolution for ideation
12 IDEAgent arXiv'26 LLM Internal Quality-diversity search over the research idea space
External Signal-Driven Generation
13 MOOSE-Chem ICLR'25 External Signal Rediscovers hypotheses from 51 high-impact papers.
14 Nova arXiv'24 External Signal 3.4× more novel ideas; 2.5× more top-rated.
15 SciAgents arXiv'24 External Signal Multi-agent reasoning over scientific knowledge graphs.
16 SciPIP arXiv'24 External Signal Multi-domain; paper-anchored idea generation.
17 IdeaSynth CHI'25 External Signal 20-user study; more alternatives explored vs. baseline.
18 MOOSE-Chem2 NeurIPS'25 External Signal Fine-grained, experimentally actionable hypotheses.
19 FlowPIE arXiv'26 External Signal Higher novelty, feasibility, and diversity vs. baselines.
20 HALO arXiv'26 External Signal Interactive co-abductive reasoning for hypothesis generation
21 TCA-SIR arXiv'26 External Signal Target-conditioned abstractions for inspiration retrieval
22 ECLAIR arXiv'26 External Signal Causally-grounded discovery in empirical software engineering
Multi-Agent Collaborative Generation
23 Combi. Creativity arXiv'24 Multi-Agent +7–10% similarity scores; cross-domain composition.
24 Deep Ideation arXiv'25 Multi-Agent +10.67% quality; surpasses conference acceptance levels.
25 VirSci ACL'25 Multi-Agent Outperforms single-agent on novelty.
26 Multi-Agent Dial. SIGDIAL'25 Multi-Agent Optimal at 3 critique-revision rounds.
27 Artificial Hivemind NeurIPS'25 Multi-Agent 26K queries; diversity collapse across models.
28 Auditable AI Sci. arXiv'26 Multi-Agent Hypothesis evolution protocol for auditable LLM scientists.
29 Diverse Personalized Ideation arXiv'26 Multi-Agent Counters AI-induced homogenization across users
Novelty and Feasibility Assessment
30 IdeaBench KDD'25 Evaluation 2,374 papers; 8 domains; novelty >0.6, feasibility <0.5.
31 LiveIdeaBench arXiv'24 Evaluation 40+ models; 1,180 keywords; 22 scientific domains.
32 AI Idea Bench 2025 arXiv'25 Evaluation 3,495 papers; alignment + general reference evaluation.
33 HeurekaBench ICLR'26 Evaluation +22% with critic module; open-ended science tasks.
34 ResearchBench ACL'26 Evaluation 12 disciplines; inspiration retrieval + ranking.
35 HindSight arXiv'26 Evaluation LLM novelty negatively correlated with impact (ρ=−0.29).
36 Diverse Hypo. Search arXiv'26 LLM Internal Diversity-aware population search; molecular/equation/algorithm discovery
37 SoundnessBench arXiv'26 Evaluation Tests if AI scientists distinguish sound vs. flawed ideas; optimism bias
38 LLM-Judge Novelty arXiv'26 Evaluation Limits of LLM-as-Judge for novelty assessment
39 LigBench arXiv'26 Evaluation Human-aligned benchmark for LLM research idea generation
40 Reconstruction arXiv'26 Evaluation Blind recovery of ideas from pre-publication bibliographies
S2 Literature Review 58 methods Expand
# Method Venue Category Evaluation Links
Literature Retrieval
1 CiteME arXiv'24 Retrieval Citation fidelity benchmark.
2 LitLLM arXiv'24 Retrieval LLM + academic database integration.
3 LitSearch arXiv'24 Retrieval Retrieval precision benchmark.
4 PaperQA2 arXiv'24 Retrieval Matches/exceeds expert on 3 tasks; 70% contradiction validation.
5 OpenResearcher EMNLP'24 Retrieval RAG + graph traversal for literature exploration.
6 PaSa arXiv'25 Retrieval Agentic multi-step iterative retrieval.
7 Search, Inspect, Fetch arXiv'26 Retrieval Structure-aware Boolean retrieval for deep-research agents
8 Rubric Reranker arXiv'26 Retrieval Trains document rerankers against search rubrics
9 Personalized DR Refinement arXiv'26 Retrieval Graph-scaffolded evidence grounding for query refinement
Survey & Related Work Generation
10 ChatPaper GitHub'23 Generation 19K+ GitHub stars; arXiv summarization tool.
11 PaperQA arXiv'23 Generation 8K+ GitHub stars; RAG for scientific Q&A.
12 AutoSurvey arXiv'24 Generation First end-to-end LLM survey drafting system.
13 GPT Researcher GitHub'24 Generation 26K+ GitHub stars; comprehensive report generation.
14 LLMs for Lit. Review arXiv'24 Generation Hallucination analysis; models still generate errors.
15 STORM arXiv'24 Generation Multi-perspective question-asking for outlines.
16 Agentic AutoSurvey arXiv'25 Generation Multi-agent role decomposition.
17 Citegeist arXiv'25 Generation Dynamic RAG pipeline on arXiv corpus.
18 IterSurvey arXiv'25 Generation Iterative outline planning with stability checks.
19 LiRA arXiv'25 Generation Multi-agent retrieval + verification + narrative.
20 SurveyForge arXiv'25 Generation Outperforms AutoSurvey on outline quality.
21 SurveyG arXiv'25 Generation Three-layer citation graph (Foundation/Dev/Frontier).
22 SurveyX arXiv'25 Generation +0.259 content quality improvement; near expert level.
23 InteractiveSurvey arXiv'25 Generation User-customizable reference categorization + outlines.
24 CiteLLM arXiv'26 Generation Hallucination-free via trusted repository routing.
25 STRUCTSURVEY arXiv'26 Generation Structured agentic retrieval for survey generation.
Deep Research Agents
26 ASReview Nature MI'21 Deep Research Active learning; up to 95% effort reduction.
27 CHIME arXiv'24 Deep Research Hierarchical organization of scientific studies.
28 DeepResearch-Agent GitHub'25 Deep Research Hierarchical multi-agent; planner + sub-agents.
29 DeerFlow GitHub'25 Deep Research Sub-agents with shared memory; sandboxed execution.
30 OpenScholar Nature'26 Deep Research 45M papers; +6.1% over GPT-4o, +5.5% over PaperQA2.
31 AutoAgent arXiv'25 Deep Research Universal LLM compatibility; GAIA benchmark.
32 Tongyi DeepResearch GitHub'25 Deep Research 30.5B params (3.3B activated); SOTA on Deep Research.
33 O-Researcher arXiv'26 Deep Research Multi-agent distillation + agentic RL.
34 OpenResearcher (2026) arXiv'26 Deep Research 54.8% BrowseComp-Plus; 97K+ trajectories.
35 AREX arXiv'26 Deep Research Recursively self-improving deep research agent
36 Predictive Navigation arXiv'26 Deep Research Pretraining objective tailored to research navigation
37 On-Device DR (4B) arXiv'26 Deep Research Exposure bounds faithfulness; retrieval bounds coverage
38 Carnot VLDB'26 Deep Research Interpretable, interactive, optimized deep-research execution
39 Marginal Value Est. arXiv'26 Deep Research Stops search when added tokens no longer change the answer
40 Retrieval-Aware Control arXiv'26 Deep Research Repairs stagnation in deep research reasoning
41 Analogical Deep Research arXiv'26 Deep Research Retrieves historical analogies for foresight analysis
42 Plato-Bio arXiv'26 Deep Research Verification-first biological novelty screening
43 Albilich arXiv'26 Deep Research Proof-state orchestration with CAS integration
Retrieval and Synthesis Quality Assessment
44 DeepScholar-Bench arXiv'25 Evaluation Coverage, coherence, factual accuracy benchmark.
45 ReportBench arXiv'25 Evaluation 100-prompt benchmark from 678 filtered survey papers.
46 IDRBench arXiv'26 Evaluation 100 tasks; interactive Deep Research evaluation.
47 ScholarGym arXiv'26 Evaluation 2,536 queries; query planning + tool invocation.
48 SciNetBench arXiv'26 Evaluation 18M papers; relation-aware retrieval <20%.
49 Self-Evolving Retrieval arXiv'26 Retrieval Agent that adapts its own literature-search policy over time
50 MasterSet arXiv'26 Retrieval Large-scale must-cite citation recommendation benchmark
51 DeepSurvey arXiv'26 Generation Evidence-constrained citation assignment; analytical depth
52 AutoResearchBench arXiv'26 Evaluation Benchmark for complex multi-hop scientific literature discovery
53 PaperMind arXiv'26 Evaluation Multimodal reasoning + critique over papers; rebuttal concerns
54 DRNOISE arXiv'26 Evaluation Benchmarks agents in misleading evidence environments
55 HiEviDR-Bench arXiv'26 Evaluation Hierarchical evidence aggregation in deep research
56 SciExplore arXiv'26 Evaluation Scientific navigation through information integration
57 WANDR arXiv'26 Evaluation Benchmark spanning wide as well as deep research
58 QA-to-DR Bench arXiv'26 Evaluation Verifiable tasks grown by iterative task evolution
S3 Coding & Experiments 45 methods Expand
# Method Venue Category Evaluation Links
Code Generation
1 SWE-bench ICLR'24 Code Gen. 2,294 real GitHub issues; Verified split (500 problems).
2 SWE-agent arXiv'24 Code Gen. Agent–computer interface paradigm for coding.
3 OpenHands ICLR'25 Code Gen. Open platform for generalist coding agents.
4 SWE-bench Pro arXiv'25 Code Gen. 1,865 enterprise problems; best score 23%.
5 SWE-EVO arXiv'25 Code Gen. Software evolution benchmark; best score 25%.
Paper-to-Code
6 FunSearch Nature'24 Paper-to-Code New cap-set solutions; evolutionary program search.
7 SciCode arXiv'24 Paper-to-Code Research-level coding across math, physics, chemistry.
8 PaperBench arXiv'25 Paper-to-Code 20 ICML'24 papers; 8,316 gradable subtasks.
9 PaperCoder arXiv'25 Paper-to-Code 3-stage multi-agent; ML papers to code repos.
10 ResearchCodeBench arXiv'25 Paper-to-Code 212 novel ML tasks; best 37.3% (Gemini-2.5-Pro).
11 SciReplicate-Bench arXiv'25 Paper-to-Code 100 tasks from 36 NLP papers; 39% ceiling.
Experiment Execution & Orchestration
12 BioPlanner arXiv'23 Execution Biological protocol planning evaluation.
13 CRISPR-GPT arXiv'24 Execution Gene-editing experiment design assistance.
14 DS-Agent arXiv'24 Execution End-to-end data science workflow automation.
15 MLE-Bench arXiv'24 Execution 75 Kaggle competitions benchmark.
16 MLAgentBench arXiv'24 Execution 13 ML experimentation tasks benchmark.
17 MLR-Copilot arXiv'24 Execution IdeaAgent + ExperimentAgent dual-agent pipeline.
18 AIDE arXiv'25 Execution SOTA on MLE-Bench + RE-Bench; tree search in code space.
19 AlphaEvolve arXiv'25 Execution LLM-generated mutations + automated evaluators.
20 AutoReproduce arXiv'25 Execution Paper lineage algorithm for experiment reproduction.
21 CURIE arXiv'25 Execution Rigorous automated experimentation framework.
22 MLGym arXiv'25 Execution AI research agent gym benchmark.
23 MLR-Bench arXiv'25 Execution 201 tasks (NeurIPS/ICLR/ICML); 80% fabrication rate.
24 Execution-Grounded arXiv'26 Execution 69.4% vs 48.0% GRPO; parallel GPU search.
25 Learn to Discover arXiv'26 Execution Test-time training + RL; math, GPU kernel, biology.
26 SciNav arXiv'26 Execution Pairwise tree-search branch selection.
27 FrontierScience arXiv'26 Execution Expert-level tasks; Olympiad + PhD difficulty.
28 AutoTTS arXiv'26 Execution Coding-agent discovery of test-time scaling strategies
29 AutoScientists arXiv'26 Execution Self-organizing agent teams for long-running experiments.
30 EurekAgent arXiv'26 Execution Agent environment engineering for autonomous discovery.
Code Correctness and Reproducibility Assessment
31 DiscoveryBench arXiv'24 Analysis Data-driven insight extraction benchmark.
32 DiscoveryWorld arXiv'24 Analysis 120 tasks; 8 topics; 3 difficulty levels.
33 InfiAgent-DABench arXiv'24 Analysis End-to-end data analysis workflow benchmark.
34 ScienceAgentBench arXiv'24 Analysis Rigorous data-driven scientific discovery assessment.
35 LAB-Bench arXiv'24 Execution Multi-domain biology research task benchmark.
36 KernelBench arXiv'25 Execution GPU kernel generation benchmark.
37 TritonBench arXiv'25 Execution Triton operator generation benchmark.
38 AstaBench arXiv'25 Execution 2,400+ problems; multi-domain scientific research.
39 ResearchClawBench arXiv'25 Execution Scientist-aligned workflow benchmark.
40 EXP-Bench ICLR'26 Execution 461 tasks from 51 AI papers.
41 PostTrainBench arXiv'26 Execution LLM post-training automation benchmark.
42 EvoDS arXiv'26 Execution Self-evolving data-science agent; skill learning + context mgmt
43 MLReplicate arXiv'26 Analysis End-to-end ML reproducibility benchmark for autonomous systems
44 BeyondSWE arXiv'26 Execution 500 instances / 246 repos; cross-repo, domain-fix, migration, doc2repo
45 NatureBench arXiv'26 Benchmark Can coding agents match Nature-family paper SOTA?
S4 Tables & Figures 38 methods Expand
# Method Venue Category Evaluation Links
Scientific Figure Generation
1 ChartGPT arXiv'23 Data Viz 6-step reasoning for chart generation.
2 MatPlotAgent arXiv'24 Data Viz +12.3 over GPT-4 base; VLM visual feedback.
3 CoDA arXiv'25 Data Viz +41.5% over baselines; multi-agent collaboration.
4 PlotGen arXiv'25 Data Viz 4–6% improvement over baselines.
5 VIS-Shepherd arXiv'25 Figure Editing Constructive critique feedback framework.
6 DiagramAgent CVPR'25 Data Viz 4 specialized agents; 8 diagram categories.
7 StarVector CVPR'25 Method Diagrams Scalable SVG generation from descriptions.
8 VisCoder EMNLP'25 Data Viz VisCode-200K dataset; 90%+ execution pass rate.
9 AI-Generated Figures arXiv'26 Policy Publisher policy survey (Nature, Science, etc.).
10 AutoFigure-Edit arXiv'26 Method Diagrams Editable text-to-SVG scientific illustrations.
11 AutoFigure ICLR'26 Method Diagrams FigureBench (3,300 pairs); publication-ready illustrations.
12 PaperBanana arXiv'26 Method Diagrams 292 test cases; outperforms baselines.
13 SAIL arXiv'26 Figure Editing Domain logic / code syntax separation.
14 Can AI Draw Sci. arXiv'26 Benchmark Benchmark for scientific figure generation by T2I/MLLMs.
15 SciDiagramEdit arXiv'26 Figure Editing Learns diagram edits from paper revision histories
16 GenGA arXiv'26 Figure Editing Editable, data-grounded graphical abstract generation
Table Understanding & Generation
17 ArxivDIGESTables EMNLP'24 Table Gen. Literature comparison table synthesis.
18 Chain-of-Table ICLR'24 Table Reasoning Multi-step table reasoning chains.
19 ShowTable CVPR'26 Table Viz Collaborative reflection and refinement.
20 Table2LaTeX-RL arXiv'25 Table Conversion Image-to-LaTeX via reinforced multimodal LM.
Mathematical Formulas & TikZ
21 AutomaTikZ ICLR'24 TikZ DaTikZ: first large-scale dataset (120K drawings).
22 DeTikZify NeurIPS'24 TikZ 360K TikZ graphics; MCTS iterative refinement.
23 TikZilla arXiv'26 TikZ 3B/8B matches GPT-5; SFT+RL on expanded DaTikZ.
24 Edit2TikZ arXiv'26 TikZ Benchmark for instruction-conditioned TikZ figure editing
Visual Fidelity and Scientific Accuracy Assessment
25 PlotCraft arXiv'25 Benchmark 1K-task benchmark; 48 chart types.
26 TeXpert SDP'25 Benchmark 3-level difficulty; 78.8%/58.7%/17.5%.
27 AbGen ACL'25 Benchmark 1,500 ablation studies; 807 NLP papers.
28 SciFig arXiv'26 Benchmark Rubric-based evaluation; 2K+ pipeline figures.
29 SciFlow-Bench arXiv'26 Benchmark 500 framework figures; inverse-parsing evaluation.
30 FigureBench ICLR'26 Benchmark 3,300 text-figure pairs; publication-ready evaluation.
31 Crafter arXiv'26 Method Diagrams Multi-agent editable figure generation from heterogeneous inputs
32 DiagramRAG arXiv'26 Method Diagrams Retrieval-augmented scientific diagram generation
33 GeoSVG-RL arXiv'26 Method Diagrams Geometry-aware RL for layout-constrained text-to-SVG diagrams
34 CSPO arXiv'26 Table Conversion Component-specific optimization for table-to-LaTeX generation
35 SciFigQual-Bench arXiv'26 Benchmark Figure quality scored with full-manuscript context
36 SciFigAlign arXiv'26 Benchmark Aligns figures with manuscript evidence for scoring
37 SciFigPlag-Bench arXiv'26 Benchmark Provenance-aware scientific figure plagiarism detection
38 VLM Blind/Misled arXiv'26 Benchmark Behavioral evaluation of VLMs on scientific figures
S5 Paper Writing 31 methods Expand
# Method Venue Category Evaluation Links
Semi-Automated Writing Assistance
1 CoAuthor arXiv'22 Collaborative Human–AI collaborative writing workflows.
2 Script&Shift CHI'25 Source Transform CHI Honorable Mention; preserves cognitive engagement.
3 AI Writing Study AIED'25 Empirical Study 90-student RCT; purposeful AI fosters writing.
4 OpenDraft GitHub'25 Full Draft Gen. 19 agents; 20K+ words in 10 min; verified citations.
5 DraftMarks arXiv'25 Transparency Skeuomorphic visual traces for AI process transparency.
6 PaperDebugger arXiv'25 In-Editor Assist Multi-agent Overleaf plugin (Reviewer+Enhancer+Scorer).
7 ScholarCopilot arXiv'25 Citation Assist 40.1% top-1 citation accuracy (vs 15.0% E5-Mistral).
8 XtraGPT arXiv'25 Post-Writing 1.5B–14B models; 7K papers; 140K revision pairs.
9 LimAgents arXiv'26 Limitations Gen. OpenReview comments + citation network integration.
10 AutoSupervision arXiv'26 In-Editor Assist Grounded revision verification closes the feedback loop
11 ReasFlow arXiv'26 In-Editor Assist Knowledge-structured assistance for applied mathematics
Fully Automated Paper Generation
12 CycleResearcher ICLR'25 E2E Gen. 5.36 ICLR scale (vs 5.24 preprint, 5.69 accepted).
13 Agent Laboratory EMNLP'25 E2E Gen. $2–13/paper; 84% cost reduction; 3.5–4.0 score.
14 FutureGen arXiv'25 Section Gen. RAG-based Future Work section generation.
15 AI Scientist Nature'26 E2E Gen. $15/paper; end-to-end across 3 ML subfields.
16 APRES arXiv'26 Rubric Revision 79% expert preference; citation-predictive rubrics.
17 RWGBench arXiv'26 Section Gen. Evaluates scholarly positioning in related-work generation.
Societal Analysis
18 AI Writing Adoption Nature'26 Measurement 41.3M papers; AI expands impact but contracts focus.
19 Nature AI Survey Nature'26 Survey 57% of researchers use AI in peer review.
20 Denial of Science arXiv'26 Measurement Indirect data poisoning could industrialize scientific fraud.
21 AI Slop OSS arXiv'26 Measurement AI-generated contributions strain open-source sustainability.
Writing Quality and AI Detection Assessment
22 Mapping LLM Use arXiv'24 Detection Up to 17.5% of CS papers AI-modified.
23 CycleReviewer ICLR'25 AI Judge 26.89% MAE reduction vs individual human reviewers.
24 Stanford Agentic Web'25 AI Judge ρ=0.42 vs human ρ=0.41; matches consistency.
25 SciIG arXiv'25 Writing Bench NAACL/ICLR 2025 introduction writing benchmark.
26 Watermarking arXiv'25 Detection Near-zero false-positive rate under controlled conditions.
27 PaperWritingBench arXiv'26 Benchmark 200 reverse-engineered top-tier conference papers.
28 PaperMentor arXiv'26 In-Editor Assist Human-centered multi-agent Overleaf writing tutor
29 LECTOR arXiv'26 Section Gen. Introduction generation via joint reasoning-graph optimization
30 CiteTracer arXiv'26 Detection Multi-agent citation-hallucination detection
31 Process Eval arXiv'26 Writing Bench Keystroke-level study of AI- vs. human-draft revision
S6 Peer Review 48 methods Expand
# Method Venue Category Evaluation Links
Automated Review Generation
1 ChatReviewer GitHub'23 Review Gen. ChatGPT-based strengths/weaknesses analysis.
2 AI-Peer-Review GitHub'24 Review Gen. Multi-LLM independent reviews + meta-review synthesis.
3 MARG arXiv'24 Review Gen. 3.7 good comments/paper (2.2× over baseline).
4 Reviewer2 arXiv'24 Review Gen. Two-stage prompt-based review aspect modeling.
5 ReviewRL EMNLP'25 Review Gen. RL + RAG; factually grounded reviews.
6 DeepReviewer arXiv'25 Review Gen. 88.21% win rate vs GPT-o1; 64% accept/reject acc.
7 OpenReviewer NAACL'25 Review Gen. Llama-8B fine-tuned on 79K expert reviews.
8 REMOR arXiv'25 Review Gen. Multi-objective RL review optimization.
9 ScholarPeer arXiv'26 Review Gen. Context-aware multi-agent; literature verification.
10 PeerCheck arXiv'26 Review Gen. Pushes LLM-generated reviews toward human-level quality.
11 Local Pre-Screening arXiv'26 Review Gen. On-premise pre-screening before human triple-blind review
12 ReVoicer arXiv'26 Review Gen. Conversational voice annotation for LLM-assisted review
Meta-Review & Reviewer Matching
13 AgentReview EMNLP'24 Meta-Review Full review lifecycle simulation; social/authority bias.
14 Meta-Review LLMs NAACL'25 Meta-Review 40 ICLR papers; GPT-3.5/LLaMA-2/PaLM-2 compared.
15 RATE arXiv'26 Matching Expertise-based matching via profile distillation.
Adversarial Attacks & Bias Analysis
16 Raina et al. EMNLP'24 Adversarial Benign adjectives as universal adversarial triggers.
17 AI Review Lottery arXiv'24 Bias Analysis 15.8% ICLR reviews AI-assisted; +4.9pp borderline.
18 Ye et al. arXiv'24 Adversarial Scores inflated to ~8; 5% manipulation flips 12%.
19 Breaking the Reviewer arXiv'25 Adversarial Systematic adversarial robustness evaluation.
20 LLM Reviewer Bias arXiv'25 Bias Analysis 1,441 papers; 95.8% rejected misclassified as acceptable.
21 Prompt Injection arXiv'25 Adversarial White-text injection; up to 100% acceptance scores.
22 Sahoo et al. arXiv'25 Adversarial +13.95 on Mistral; 13 LLMs; 15 attack strategies.
23 Zhou et al. arXiv'25 Adversarial +1.24 to +2.80 from hype prose; 10.00 under iteration.
24 Gaming AI Reviews arXiv'26 Adversarial Gaming AI-assisted peer reviews poses new risks.
25 Phantom Refs arXiv'26 Adversarial Hallucinated citations survive top-tier peer review.
26 Rhetorical Reward-Hacking arXiv'26 Adversarial Confident framing shifts AI review independent of substance
Detection & Policy
27 AI Detection arXiv'25 Detection 788,984 AI-written reviews; 18 detection algorithms.
28 AI Use Rejects Nature'26 Policy 497 papers rejected (~2% of submissions).
29 Nature AI Survey Nature'26 Survey 1,600 academics; 57% use AI in peer review.
30 Policy Enforcement arXiv'26 Policy All 5 SOTA detectors misclassify LLM-polished reviews.
31 Reviewer Feedback CHI'26 Empirical ICLR 2025 live process; reviewer engagement study.
32 AAAI-26 Pilot arXiv'26 Empirical AI-assisted peer review at scale; the AAAI-26 pilot.
33 Reviewer AI Policies arXiv'26 Policy Maps community AI policies against resulting review quality
Review Consistency and Bias Assessment
34 Review Survey IF'25 Survey Comprehensive taxonomy of review methods.
35 Stanford Agentic Web'25 Quality ρ=0.42 vs human ρ=0.41; matches consistency.
36 ClaimCheck EMNLP'25 Quality LLM critique grounding; gaps in factual basis.
37 ReViewGraph AAAI'26 Quality +15.73% avg improvement via heterogeneous graph.
38 ReviewAgents arXiv'25 Quality 37,403 papers; 142,324 reviews; Review-CoT dataset.
39 ICLR 2025 Study NMI'26 Quality 22,467 reviews; 89% quality improved; no acceptance effect.
40 AI Reviewer Limits arXiv'26 Quality 45 experts assess AI vs. human reviews of Nature-family papers
41 PRISM arXiv'26 Quality Multi-dimensional benchmark for LLM peer reviewers
42 ProReviewer arXiv'26 Review Gen. Proactive investigation agent; probes suspicious paper parts
43 MERIT arXiv'26 Matching Rubric-informed reviewer assignment via expertise modeling
44 Presentation Gaming arXiv'26 Adversarial Presentation-only revisions game AI review without prompts
45 LLMs Favor LLMs? arXiv'26 Bias Analysis Apparent self-preference is largely general leniency
46 LLM-Human Alignment arXiv'26 Quality Measures alignment of LLM reviews with human peer review
47 Epistemic Reliability arXiv'26 Quality Tests genuine analysis vs articulated intuition in judges
48 SurveyReview arXiv'26 Quality Reviewer-aligned benchmark for survey evaluators
S7 Rebuttal & Revision 15 methods Expand
# Method Venue Category Evaluation Links
Reviewer Comment Analysis
1 ReviewMT arXiv'24 Analysis 26,841 papers; 92,017 reviews; multi-turn dialogue.
2 ICLR Rebuttal Study arXiv'25 Analysis ICLR 2024–2025; score transition analysis.
Automated Rebuttal Generation
3 ReviewerToo arXiv'25 Modular Pipeline 81.8% accept/reject accuracy (vs 83.9% human).
4 RebuttalAgent ICLR'26 Rebuttal Gen. 18.3% avg improvement; ToM-grounded.
5 Author-in-the-Loop ACL'26 Author-Aware Integrates author expertise and intent.
6 DRPG arXiv'26 Rebuttal Gen. 98%+ planning accuracy; surpasses avg human quality.
7 Paper2Rebuttal arXiv'26 Rebuttal Gen. Evidence-centric rebuttal planning.
Rebuttal Effectiveness Assessment
8 Re² arXiv'25 Dataset 19,926 submissions; 70,668 reviews; 53,818 rebuttals.
9 Commitment Checklist arXiv'26 Benchmark 11.8 commitments/paper; ~25% unfulfilled.
10 Re³Align ACL'26 Dataset First large-scale aligned review–response–revision triplets.
11 RbtAct arXiv'26 Analysis Rebuttals as supervision for actionable review feedback
12 GoodPoint arXiv'26 Analysis Learns constructive feedback grounded in author responses
13 Defend arXiv'26 Rebuttal Gen. Minimal-guidance rebuttal generation; improved factual correctness
14 Rebuttals Move arXiv'26 Analysis Rebuttals move scores, bounded by initial-review structure.
15 Trust AI Reviews arXiv'26 Analysis Authors' response to AI-based reviews.
S8 Paper2X (Dissemination) 32 methods Expand
# Method Venue Category Evaluation Links
Paper2Poster
1 P2P ICLR'26 Paper2Poster P2PInstruct 30K+ examples; 3 specialized agents.
2 Paper2Poster NeurIPS'25 Paper2Poster $0.005/poster; 87% fewer tokens vs GPT-4o.
3 PosterForest arXiv'25 Paper2Poster Hierarchical multi-agent collaboration.
4 PosterGen arXiv'25 Paper2Poster Aesthetic-aware multi-agent generation.
5 APEX arXiv'26 Paper2Poster First agentic interactive poster editing.
6 PosterOmni arXiv'26 Paper2Poster 6 unified poster tasks; outperforms open-source.
Paper2Slides
7 DOC2PPT AAAI'22 Paper2Slides 5,873 paired document–slide decks.
8 PPTAgent EMNLP'25 Paper2Slides PPTEval benchmark; 10,448 curated presentations.
9 AutoPresent CVPR'25 Paper2Slides 8B Llama model; SlidesBench (7K train, 585 test).
10 Paper2Slides GitHub'25 Paper2Slides 4-stage RAG pipeline; one-click conversion.
11 Auto-Slides arXiv'25 Paper2Slides Multi-agent Beamer generation; interactive editing.
12 PASS arXiv'25 Paper2Slides First combined slides + AI audio delivery.
13 SlideGen arXiv'25 Paper2Slides Multi-agent VLM coordination; editable PPTX output.
14 Talk to Your Slides arXiv'25 Paper2Slides +34% instruction fidelity; 87% lower cost vs GUI.
15 SlideTailor AAAI'26 Paper2Slides User-preference conditioned; chain-of-speech.
16 DeepPresenter arXiv'26 Paper2Slides 9B model competitive with frontier at lower cost.
17 Office Raccoon Web'26 Paper2Slides Page-level editing; template/brand-guideline learning.
Paper2Video
18 Preacher ICCV'25 Paper2Video Top-down decomposition; 5 research fields.
19 Paper2Video arXiv'25 Paper2Video 101 paper–video pairs; +10% PresentQuiz accuracy.
20 PresentAgent EMNLP'25 Paper2Video PresentEval benchmark; approaches human-level.
21 Paper2Video Talks arXiv'26 Paper2Video A good talk teaches takeaways, not summaries.
Paper2Web & Social Media
22 Paper2Web arXiv'25 Paper2Web 10,716 papers; multimedia-rich academic homepages.
23 ResearchStudio-Reel arXiv'26 Paper2Web Automates paper to poster, video, and blog jointly.
24 I-WebGenBench arXiv'26 Paper2Web Evaluates interactivity in LLM-generated scientific web apps.
25 SciForge arXiv'26 Paper2Web AI-native multimodal workbench for discovery and presentation
Fidelity and Adoption Assessment
26 PPTEval EMNLP'25 Benchmark Content, design, coherence; 10,448 presentations.
27 PresentQuiz arXiv'25 Benchmark 101 paper–video pairs; +10% over human on comprehension.
28 PresentEval EMNLP'25 Benchmark End-to-end narrated video quality; near human-level.
29 Any2Poster arXiv'26 Paper2Poster Poster generation from eight input modalities across domains
30 X+Slides arXiv'26 Paper2Slides Audience-conditioned slide generation benchmark
31 PresentAgent-2 arXiv'26 Paper2Video Generalist agentic presentation-video generation
32 Sci. Comm. Correspondence arXiv'26 Benchmark Fine-grained paper--slides--video correspondence

Call for Contributions
Call for Contributions

This collection is a living resource. If you have a new method, benchmark, or tool that fits one of the eight stages, submit a pull request with the paper title, venue, link, and a brief evaluation note.

Submit a Pull Request



Exhibition

AI-Generated Dissemination Artifacts
Poster full view
Subsection · 03 AI-Generated Posters

Paper2Poster

A virtual gallery of research posters generated by frontier AI models for this survey. Each poster is produced end-to-end by a different system, showcasing the current state of automated scientific dissemination.

01 Gemini Pro Poster
Gemini Pro
02 Qwen Image 2.0 Poster
Qwen-Image-2.0
03 Claude Opus 4.7 Poster
Opus 4.7
04 SenseNova-U1 Poster
SenseNova-U1
05 GPT-5.5 Poster
GPT-5.5
06 Seedream 4.0 Poster
Seedream 4.0
Landscape Posters
07 Gemini Pro Landscape Poster
 Landscape Gemini Pro
08 Qwen Image 2.0 Landscape Poster
 Landscape Qwen-Image-2.0
09 Claude Opus 4.7 Landscape Poster
 Landscape Opus 4.7

Citation

BibTeX

If you find this survey or the reading list useful, please consider citing our paper: 

@article{survey-ai-auto-research,
  title   = {{AI} for {Auto-Research}: Roadmap \& User Guide},
  author  = {Kong, Lingdong and Sun, Xian and Chow, Wei
             and Li, Linfeng and Lin, Kevin Qinghong and Zhang, Xuan Billy
             and Wang, Song and Li, Rong and Wu, Qing
             and Gao, Wei and Wang, Yingshuo and Xie, Shaoyuan
             and Liu, Jiachen and Qu, Leigang and Li, Shijie
             and Ng, Lai Xing and Cottereau, Benoit R.
             and Liu, Ziwei and Chua, Tat-Seng and Ooi, Wei Tsang},
  journal = {arXiv preprint arXiv:2605.18661},
  year    = {2026}
}


Responsible Use

Scope & Caveats
AI Auto-Research Lifecycle Overview

This work is intended to inform responsible use of AI-assisted research tools, not to endorse replacing human scientific judgment with full automation. Current systems are most reliable when used to assist retrieval, drafting, coding, visualization, review support, and dissemination, while humans retain responsibility for novelty, interpretation, verification, authorship, and accountability.

Because the field evolves rapidly, this paper should be read as a structured snapshot through our search cutoff, and AI-generated research outputs should be independently verified before scholarly use.


Contributors

Lingdong Kong*
National University of Singapore
Xian Sun*
Duke University
Wei Chow*
National University of Singapore
Linfeng Li
National University of Singapore
Kevin Qinghong Lin
University of Oxford
Xuan Billy Zhang
National University of Singapore
Song Wang
Zhejiang University
Rong Li
HKUST(GZ)
Qing Wu
Nanyang Technological University, Singapore
Wei Gao
Northeastern University
Yingshuo Wang
University of California, Berkeley
Shaoyuan Xie
University of California, Irvine
Jiachen Liu
Orchestra Research
Leigang Qu
National University of Singapore
Shijie Li
A*STAR, Singapore
Lai Xing Ng
A*STAR, Singapore
Benoit R. Cottereau
CNRS & IPAL
Ziwei Liu
Nanyang Technological University, Singapore
Tat-Seng Chua
National University of Singapore
Wei Tsang Ooi
National University of Singapore

* Equal Contributions




Download PPTX