A living survey · ML / NLP / HCI
Human-Agent Collaboration in Data Science
A continuously updated collection supporting literature discovery and comparison in human-agent collaboration.
How do people and AI agents work together on data science and analytical tasks?Explore papers and benchmarks, compare selected approaches, and trace reported successes, failures, and limitations back to the original studies.
Our scope includes LLM agents for data analysis, data modeling and machine learning engineering, related graph analysis and reasoning, and supporting methods and benchmarks. The collection focuses on published research, with selected preprints.
Conference Coverage
This table defines the conference and year coverage behind the review. Each cell reports the metadata currently held in the main corpus. Partial and missing cells identify gaps that remain to be filled, rather than implying a complete survey of the field.
| Year | ICML | ACL | ICLR | EMNLP | NeurIPS | COLM | CHI |
|---|---|---|---|---|---|---|---|
| 2023 | Scanned1,828 records | Scanned1,976 records | Scanned3,810 records | Scanned2,106 records | Scanned3,394 records | N/AFirst held in 2024 | Scanned879 records |
| 2024 | Scanned2,610 records | Scanned1,915 records | Scanned5,691 records | Scanned2,271 records | Scanned4,236 records | Scanned299 records | Scanned1,057 records |
| 2025 | Scanned3,419 records | Scanned3,086 records | Scanned8,613 records | Scanned3,214 records | Scanned5,540 records | Scanned418 records | Scanned1,249 records |
| 2026 | Scanned6,555 records | Scanned4,459 records | Scanned13,715 records | Not scannedUnavailable at collection | Not scannedUnavailable at collection | Scanned856 records | Scanned1,702 records |
84,898 metadata records in the promoted corpus snapshot. Scanned means metadata collected under the selected source and track policy, not every paper manually read. Counts include public rejected papers where collected; ACL/EMNLP include Findings; CHI covers main research articles. Missing years reflect this snapshot, not current venue availability.
The Collection at a Glance
Broad metadata coverage supports a smaller screened collection. A human-selected subset receives full-paper analysis.
- Candidate papers
- 605Keep + borderline in the main screened collection
- Deep-analysis papers*
- 90Human-selected papers with full-text paper cards
* Includes relevant preprints manually selected for deep analysis.
Full corpus dashboard →keep and borderline are LLM judgments of relevance to our research scope.
Explore the Survey
Browse papers and benchmarks, compare selected work in the Research Space, or explore reported failures through Challenges.
Papers →
Browse candidate papers, abstracts, and relevance assessments across the reviewed literature.
Evaluation resourcesBenchmarks →
Datasets, evaluation tasks, and environments introduced or extended by papers in the screened collection.
Corpus coverageCoverage →
Inspect venue and year coverage, screening counts, and the composition of the corpus.
Selected papersResearch Space →
Compare selected papers by their tasks, human involvement, and agent capabilities, with citation links within the collection.
Cross-paper analysisChallenges →
Explore recurring failures and limitations reported in papers, grouped into challenges with traceable source evidence.
External linksResources →
Useful links on human-agent collaboration and automated data science.
Recent Updates
- UI refresh and reading guides
Refreshed the Research Space, Challenges, and corpus dashboard with consistent styling and clearer navigation. Added analysis reading guides, refined the overview, and made the coverage table easier to browse. No papers were added or analysis rerun for this interface update.
- ACL Findings and COLM 2026 added
67 screened candidates and nine deep-analysis papers join the collection. The research space now covers 90 papers, with 706 evidence entries across 20 challenge clusters and two entries awaiting assignment.
- 32 new benchmark and resource papers
Benchmark screening and task classification now cover the new candidates. Browse 241 resource-contributing papers by Data Analysis, Data Modeling, Mixed, or Others; five uncertain entries remain separate.
- Papers, conferences, and benchmarks
Dedicated literature and coverage directories now connect to a benchmark collection, identified from candidate-paper metadata with supporting source excerpts.
- A survey-first overview
Conference coverage and collection statistics now lead the overview. Existing paper, research-space, and challenge reports remain available. This interface update does not add papers or rerun the analysis.
About This Survey
An experimental, human-curated and LLM-assisted literature resource. Human selection guides deeper analysis of the screened collection.
Reading the Analysis
Analysis is provided for exploration, with source evidence available for inspection. Extracted evidence and synthesis have not all been independently verified.