deep-research

Deep Research conducts systematic academic literature reviews in six strictly sequential phases, producing structured notes, a curated paper database, and a synthesized final report organized by phase. It solves the problem of shallow, abstract-only literature surveys by forcing the agent through frontier discovery, a broad survey, deep reading of full papers, a code-and-tools scan, synthesis, and final report assembly — with a phase-gate protocol that requires the previous phase's output files (with minimum paper/repo counts) to exist on disk before advancing.

The workflow integrates multiple scholarly sources: a conference-paper finder, the arXiv API, the Semantic Scholar Graph API, OpenReview, and ar5iv HTML rendering, with bundled scripts for searching, downloading papers, extracting PDF text, managing BibTeX, maintaining a paper database, and compiling the report. It enforces a paper-quality policy that prioritizes peer-reviewed conference and journal work over preprints and marks preprint status explicitly. Output lands in a per-topic folder structured by phase (frontier, survey, deep dive, code, synthesis, report), and each phase prints a completion checkpoint listing the files it wrote.

Target users are researchers, graduate students, and engineers who need a rigorous, reproducible survey of a technical topic — especially in AI/ML — rather than a quick summary. It is best suited to producing a defensible state-of-the-art review grounded in papers that were actually read, with citations traced to specific findings.