Sakana AI has built a language-model system that reads research papers and hunts for the faults human reviewers are supposed to catch.
The Tokyo lab says its reviewer identified 73 percent of errors in the core claims of papers it was given, a figure that puts automated screening in the same territory as overstretched volunteer referees.
The pitch arrives at an awkward moment. arXiv is capping submissions precisely because human moderators cannot keep up with the flood, and labs are churning out papers faster than any volunteer pool can absorb.
Sakana’s approach targets the central assertion of a paper rather than surface issues like formatting or citation style, which is where most automated checks have stopped.
Independent verification is the unresolved question. A model grading claims still needs ground truth about whether a result holds, and errors of the kind that matter most often surface only when someone tries to reproduce the work.
The system also raises a symmetry problem for a field already leaning on machine assistance. Automated review plus automated writing risks a loop in which machine-generated claims are validated by machine readers.
Sakana has not said whether the tool will be released publicly or reserved for partners.
