Automated Literature Review#
- Point of contact: TBD
- Suitable for: TBD
- Keywords: autoresearch, literature review, automation
Today’s situation is scale without trust. Literature reviews now sit on hundreds of records, messy PDFs, and conflicting claims, so pure manual writing cannot keep up—yet one-shot model drafting is too free: it either drops material or fakes coverage with bulk citations. An automatic review is really a multi-stage system—collect, filter, extract claims, bind evidence, review, structure, cite, and typeset—where each stage can fail differently and a fluent paragraph can hide a broken chain of custody. The hard problem is to automate synthesis without losing traceability, and to grow the report without growing citation theater.
Our approach freezes that chain and writes by compress-then-expand. Data collection locks a corpus; intelligence analysis builds an evidence-bound claim graph, with light approval when heavy re-review is waste; report writing compresses claims into main arguments, plans subsections, expands locally with tight citation clusters, adds introduction and summary, validates dual binding, and renders PDF under a resumable workspace. Gaps remain in quality and adaptation: coverage can still be mechanical rather than deep, sections can skew, expand can wobble, and the pipeline is still a batch freeze rather than continuous auto literature review as new material arrives. Future work is to raise the report to top-level quality—faithful to the evidence, insightful in synthesis—and to accept new materials continuously, so automatic literature review stays current without sacrificing honesty.