A new arXiv paper says today's AI co-scientist tools treat every researcher the same, and proposes fixing that with personalization.
The paper, titled "Personalized Auto-Research: Towards a True AI Co-Scientist," argues that current AI co-scientist systems generate hypotheses, search related work, run experiments, and draft papers while ignoring who actually asked. Two researchers with the same goal but different backgrounds, methods, and collaborators currently get back nearly identical output, a pattern the authors call a one-size-fits-all failure mode. Their proposed framework instead builds a graph-grounded profile of each researcher and threads it through every stage of the pipeline, from literature retrieval to hypothesis generation to peer review. The paper does not ship a working system; it lays out the framework and flags open problems still to solve.
Research taste is not generic. What counts as novel or feasible depends on a scientist's prior work and the community they answer to, and a tool that erases that context risks flattening the tacit knowledge that produces genuinely new ideas. As AI co-scientist tools move from novelty to lab infrastructure, personalization could be the difference between a useful collaborator and an expensive autocomplete for grant proposals.
It is a framework paper, not a product, so the real test is whether anyone builds this and whether personalized suggestions actually beat the generic kind.