About Soba Researcher

Soba Researcher is an AI Research Assistant. Give it a title and a dataset, and it studies that data, reports what it finds, and writes the result up as a full research paper with figures, tables and references, delivered as LaTeX source and a finished PDF. It is a hobby project, built to explore two questions: how far a title and a dataset can be taken toward a complete, properly formatted study, and how much a paper improves when it has to answer a large panel of critical reviewers.

Research with your dataset

This is where Soba Researcher does real research work. When you add a link to your data, the paper becomes an analysis of that data and not an essay about the topic. The file is downloaded and read in full, whether it is a CSV, TSV, JSON, XLSX or plain text file, a Google Sheet or a GitHub link, up to 200,000 rows and 60 columns. Every variable is then measured. For numeric columns it computes the count, the missing values, the minimum, maximum, mean, median and standard deviation. For categorical columns it counts each category, and free-text columns such as survey answers or interview responses are read as text.

The findings in the paper come from those measurements. The sample description, the counts, the charts and the quotes are taken from your file, the descriptive charts plot real values from the data, and the sample table reports the dataset's actual composition. The methodology section describes what was actually done, a secondary analysis of your dataset with its source, its number of records and the variables used, and it does not claim interviews or fieldwork that never happened. Your dataset is cited in the reference list with its link so a reader can trace every number back to it, and a short note from you on what the data is and how it was collected is used to frame the study correctly. The exact data profile the paper was written from is shown on the paper's page, so you can check the paper against it.

Statistical tests and regression

Soba Researcher is designed to go beyond description. Statistical testing and regression are the next stage of its dataset analysis: correlations between numeric variables, comparisons between groups, tests of association between categories, and regression models that estimate how an outcome changes with its predictors, each reported with its test statistic, significance level and effect size. This stage is in development. Until it is released, the analysis in a paper is descriptive, built on the counts, means, medians and spreads measured from your data.

Experimental by design

Soba Researcher is itself an experiment in automated research. Each paper is treated as a hypothesis to be tested: it is drafted, attacked by reviewers with opposing mindsets, and rewritten, and the rewrite is kept only if it holds up against the first draft. The methods it uses are still being tried and changed, so results will vary from run to run and should be read with that in mind.

What it does

It writes the whole paper: abstract, introduction, literature review, research questions, methodology, findings, discussion, implications, limitations, future research, conclusion and a reference list. Each paper carries five figures and five tables that it draws itself in LaTeX, so they stay sharp in the PDF. You choose between a two-column IEEE conference layout and a single-column layout, and you can add up to six authors with their roles and affiliations. The LaTeX is compiled into a PDF on your own computer, and common LaTeX slips are repaired automatically so that one stray character does not cost you the PDF.

The AI Review Agents

With Review switched on, the first draft is not the final paper. It goes to a panel of 120 AI Review Agents, each with a different mindset and a different focus. Industry research styles such as Google Research, Google DeepMind, Meta FAIR, Tesla Autopilot, OpenAI, Microsoft Research and NVIDIA Research sit beside a journal editor, a methodologist, a statistician, a domain expert, an ethics and policy reviewer, a startup founder, a skeptical practitioner and a general reader. Each reviewer looks at one thing only, such as novelty, method, evidence, clarity, figures and tables, references, limitations, practical impact, structure or discussion.

A meta-reviewer then merges every review into a single decision letter, with the most-requested changes first, and the paper is written again to answer it. If the rewrite comes back shorter or loses figures or tables, the first draft is kept. The average score, the verdict split, the score from each mindset and every individual review can be read on the paper's Review tab. The reviewers are fictional personas. They are not human, and they are not affiliated with or speaking for any company named in a mindset.

Learning from real peer reviews

Soba Researcher's writing rules come from a study of 2,378 public peer reviews on OpenReview, from the ICLR 2025 conference. We looked at what reviewers criticised most often, especially in papers that were rejected. The most common criticisms were unclear writing and undefined terms, weak comparison with earlier studies, figures and tables that do not explain themselves, outdated literature, unjustified design choices, an unclear contribution, and claims stronger than the evidence.

Every paper Soba Researcher writes follows rules built to avoid these mistakes, and its panel of AI Review Agents checks for the same problems before the paper is revised. No tool can guarantee acceptance, but a draft that avoids the most common criticisms gives you a stronger starting point.

Who it is for

Soba Researcher is for students, researchers and writers who have data or a question and want a structured starting point: a study that is already analysed, written up, formatted and criticised, so the real work can start from a draft instead of a blank page.

How it works with evidence

Soba Researcher builds every study on the evidence you bring to it. Your dataset is its field: it reads the records you have collected, whether they are survey responses, interview transcripts, measurements or logs, and it describes, summarises and interprets them into findings, charts and tables. When you start from a title alone, it writes a complete model study that shows the design, the structure and the kind of results such a project would report, ready for you to replace with your own material. Its references are drafted alongside the text and your own dataset is always cited with its link. Every paper it produces is a working draft, so check it against your data and your sources, and follow your institution's or publisher's rules on the use of AI before you use it anywhere.

Developed by Kamber's Child

Soba Researcher is developed by Kamber's Child, a group working across technology, education, research and business innovation, which also runs the Kamber School of Machine Learning and Kamber Business School.

License

The site content and sample papers are shared under CC BY 4.0. For details see Copyright and License, the Terms and Conditions and the Privacy Policy.