Gemini-backed Paper Assistant Tool provides automated feedback for theoretical computer scientists at STOC 2026

Theoretical computer science runs on proofs, and proofs are unforgiving. A single miscalculated inequality or mislabeled variable can sink months of work, and normally nobody finds out until peer review, long after the damage is done. Google Research decided to see if a specialized Gemini tool could catch those mistakes earlier, so it built something called Paper Assistant Tool, or PAT, and offered it to authors submitting to STOC 2026, one of the field's most selective conferences.

PAT isn't a generic chatbot bolted onto a paper. It runs on an advanced version of Gemini 2.5 Deep Think, using inference-scaling techniques that let it generate and cross-check multiple reasoning paths instead of committing to one linear argument. That matters in math, where a single wrong turn early in a proof invalidates everything downstream. Google says this approach cuts down on hallucinated errors and helps the model zero in on the issues that actually matter, then hands authors a structured writeup: a summary of the paper's contributions, flagged problems in specific lemmas or theorems, and a list of typos and small fixes.

The results, by Google's own account, were striking. Turnaround was about two days. Over 100 authors responded to a post-experiment survey, and more than 80% of eligible papers opted into the AI review once word got around. One researcher described PAT catching a bug that had quietly broken their proof for months, calling it embarrassingly simple in hindsight — the kind of thing that's invisible until someone, or something, points it out. Ninety-seven percent said the feedback was useful and that they'd use PAT again; 81% said it made their paper clearer.

What's notable is how authors actually used the tool. They didn't treat PAT's output as gospel. Experts in the field could tell the difference between a genuine catch and an AI hallucination, especially when PAT stumbled over dense notation or figures, which it apparently still does. So the workflow that emerged was less "trust the machine" and more "use the machine as a fast first pass, then verify by hand." That's a meaningfully different posture than most AI-assistance hype, which tends to promise autonomy rather than collaboration.

Google is careful to frame this as augmentation, not a replacement for peer review, and the survey backs that framing — 88% of participants wanted continuous access to a tool like this throughout their research process, and 75% saw educational value for students learning to tighten their proofs. Whether PAT scales beyond a single conference pilot, or whether other venues start expecting AI pre-checks as a submission norm, is the more interesting question left hanging here.