Undermind

What is Undermind?

Undermind is a search agent for scientific literature that inverts the usual speed trade-off. Instead of returning results in under a second, it spends three to six minutes on a single query: asking clarifying follow-up questions to pin down what you actually mean, then scanning hundreds of papers, classifying each candidate as highly relevant, closely related, or ignorable, following citation trails outward, and adapting its strategy as it learns the shape of the topic. The underlying corpus is a Semantic Scholar index of over 200 million articles spanning PubMed, arXiv, and commercial publishers, so coverage is genuinely cross-disciplinary rather than confined to one field.

The design target is recall on hard questions rather than convenience on easy ones. Undermind’s own whitepaper — a vendor-run benchmark on its own product, not an independent evaluation — compared its v1 engine against Google Scholar across roughly 300 user queries and reported ten times more relevant results for the median query. That framing is worth taking at face value only as far as vendor benchmarks go, but it does describe the intended use case accurately: the tool earns its keep when you suspect relevant work exists and conventional keyword search is not surfacing it. On the free tier the agent reads abstracts and metadata; paid tiers allocate more compute to also read the full texts of nearly all open access and preprint articles, which is where the relevance classification gets meaningfully sharper.

Who is it for?

Undermind fits researchers whose cost of missing a paper is high — a failed grant application, a duplicated experiment, a patent filed over existing prior art. It is a poor fit for anyone wanting quick answers, general web research, or writing assistance, since it does only literature discovery and does it slowly. The pricing split matters when choosing: the widely quoted $16/month figure is the academic annual rate, and commercial users pay roughly four times that. Anyone evaluating it should also compare against Elicit and Consensus for faster extraction across papers, SciSpace for reading and explaining individual PDFs, and Perplexity for general research that spans beyond academic sources.

• Academic researchers and PhD students running systematic literature reviews or entering an unfamiliar subfield where they do not yet know the correct search vocabulary.
• Industry R&D and biotech teams doing novelty assessment and prior-art scanning, who need the recall but should budget for the $75/month industry rate rather than the academic one.
• Grant writers and principal investigators who need defensible evidence that a proposed direction has not already been covered in the published record.
• Research librarians and evidence-synthesis specialists who can absorb the three-to-six-minute search latency into a batch workflow and value the citation-trail following over instant results.
• Not a fit for casual users or students wanting fast answers — Perplexity or Consensus return usable results in seconds for straightforward questions.