Undermind

Undermind is an academic research tool. Undermind runs deep literature searches that read full papers and follow citation trails, returning cited reports and ranked reading lists. Built by MIT physics PhDs, it targets scientists and R&D teams who need thorough coverage rather than fast snippets.

Undermind is a research assistant built around scientific literature retrieval rather than general chat. A researcher describes a project in plain language, answers a few follow-up questions, and the system dispatches a search that reads and scores hundreds of papers, then follows citation trails outward until it stops finding new material. The output is a relevance-ranked table of papers plus a report with in-line citations, so each statement can be traced back to its source. The company reports that its first search engine beat Google Scholar by a factor of ten in its own benchmark, and publishes a v2 comparison against frontier models running as web search agents. The workflow is deliberately iterative. After the first pass, users open reports, pull up full texts, extract specific details, and ask for custom comparison tables. Alerts watch an area of interest and flag newly published work, which turns the tool into an ongoing monitoring layer rather than a one-off query. The site pitches this as building command of a field over time, with a shared workspace where colleagues can collaborate on the same searches and libraries. It suits working scientists, clinical investigators, and industrial R&D groups more than casual readers. GSK reports more than a thousand of its scientists on the platform, and the testimonials skew toward people chasing obscure or highly nuanced material across chemistry, drug delivery, and physics. Pricing separates individual researchers from teams, with an enterprise tier that adds organizational login, an admin dashboard, and contractual security review. The honest comparison point is the frontier chatbot: Undermind argues that general agents with web search miss key papers, and its published evaluation puts its own deep search well ahead on recall at ten minutes. The tradeoff is time. Searches run for minutes rather than seconds, which is a poor fit for quick fact checks and a good fit for a literature review that would otherwise take days.