Literature Review Using SciSpace Agent: In-Depth Walkthrough

A research agent workflow in SciSpace is built to turn a vague topic into a structured literature review—starting with a domain-grounded research question and ending with review-ready reports (including LaTeX/PDF) that carry citations, tables, and research-gap analysis. The core idea is that strong literature reviews depend less on “asking for papers” and more on iterative control: define the domain, run multi-database searches with query diversity, extract structured data into columns, then generate review outputs that map claims back to specific sources.

The process begins with forming a research question. Instead of jumping straight into paper lists, the workflow pushes users to first understand the domain via SciSpace’s “deep preview” option. Users run targeted queries (example given: algorithmic improvements in reinforcement learning from human feedback for LLM training). Deep preview expands the search through multiple elaborated queries, pulls in citation trails and references, and returns a large, relevance-ranked set (the walkthrough cites 572 papers, with the most relevant rising to the top). The results are not just a reading list: they include follow-up questions and categorization signals, plus table-like summaries that break the field into components such as optimization efficiency, feedback mechanism robustness, and architectural adaptability. The agent also surfaces strengths and weaknesses, helping users translate domain understanding into clearer research directions.

From there, the workflow moves to a comprehensive multi-database search. Earlier SciSpace literature review tools relied heavily on SciSpace Academic, but the agent integrates additional sources such as Google Scholar, PubMed, and arXiv (archive). For each selected research question, it fires multiple queries per database—using Boolean-style prompts where appropriate (especially for Google Scholar) and applying database-specific filters (arXiv filters are emphasized for AI workflows; PubMed filters are positioned for biomedical/chemical domains). Results are combined and reranked, with the system selecting the top 100 and attaching a relevance score plus reasoning. The walkthrough stresses that query breadth matters: humans miss keywords, while the agent can generate multiple query variants across themes.

Next comes result analysis, where the system shifts from “papers” to “data.” Users can add columns to extract structured fields quickly—methods, limitations, future scope, evaluation metrics, and more—often leveraging PDFs when available. A key feature highlighted is research-gap extraction by mining limitations and future directions, then producing “extract insights” outputs that summarize challenges with tight citation grounding. The agent’s approach aims to reduce hallucinations by restricting insights to highly relevant, top-ranked papers and maintaining one-to-one citation mapping.

Finally, the workflow generates custom literature review reports with citations and flexible formatting. Outputs can be written in Markdown or LaTeX, including two-column layouts and multiple citation styles (example: IEEE-style is mentioned as a default). The walkthrough demonstrates a critical review with added columns (e.g., misalignment risks, key limitations, criticisms, proposed solutions, fundamental problems), LaTeX sectioning for accuracy, and compilation to PDF. It also shows other review types: scoping reviews (with full-text PDF download and visualizations) and systematic literature reviews using a PRISMA-style plan (research question → criteria → multi-database search → screening → deduplication → full-text extraction → final PRISMA report). The overall takeaway is a research-assistant model: iterative prompts, structured extraction, and controlled report generation that can be refined through follow-up questions when outputs need correction.