Google's 145-page doc details how researchers use Gemini for scientific discovery

dzscholar · reddit · 2026-09-16

Google published a 145-page document on how researchers use Gemini for scientific problems. Key takeaways summarized by a Reddit user:

• Adversarial reviewer: Gemini found a significant flaw in a cryptography proof that had already passed human peer review — far beyond PDF summarization.
• Cross-domain transfer: the model pulls theorems from geometry and measure theory to advance algorithmic questions, leveraging broad general knowledge.
• Human-in-the-loop division of labor: humans pick problems, verify every proof, and judge novelty; Gemini suggests ideas, spots gaps, and does heavy computation.
• Automation loops, not conversations: in some projects Gemini runs a propose-expression → write-test-code → read-errors → self-correct loop, with humans stepping in only when a promising lead appears.

The overall arc: AI-assisted research is shifting from chat toward sophisticated embedded automation loops. Full document on arXiv (2602.03837).

Original post →

More from Research

• Rhoda study: web-video pre-training scales robot foundation models for real work — vincesitzmann · 2026-09-16
• TorchLean formalizes neural networks in Lean 4, from autograd to CUDA verification — srush_nlp · 2026-09-16
• Podcast: AI agents spotted a rare disease top labs missed, frontier models beat old workflows — danielmckinn0n · 2026-09-16
• An AI that only predicts, never writes: DOOM demo pitches prediction engine over LLM paradigm — danshipper · 2026-09-16
• CheatBench Debuts to Measure How Often AI Agents Game Tasks for Rewards — ricklamers · 2026-09-16
• k-server conjecture proven true: work function algorithm achieves competitive ratio k — ctjlewis · 2026-09-16

Research channel →