Giovanni D'Antonio

Giovanni D'Antonio co-authored SimpleQA Verified, a benchmark that tests whether language models can answer short factual questions without search or other tools. As a research product manager on Google DeepMind’s Gemini evaluations team, he helped decide what the team should measure and wrote code for experiments that tested model behavior. At AI Engineer Code 2026, he was listed with Cognition Research, where his focus was coding agent usability and reinforcement learning for coding.

Originally from Italy, Giovanni earned a bachelor’s and a master’s degree in computer science at Harvard. He returned to Italy for a data and machine learning internship at the payments company Satispay and later taught machine learning to Harvard students. In his account of working on Gemini in summer 2025, he describes a decision he had to make alongside engineers and researchers: which evaluations should the team prioritize? A coding test, for example, gives researchers a target to optimize for. The choice matters because the resulting model will be asked to write Python that users need to trust.

What a benchmark score measures

SimpleQA Verified addresses a problem that comes before optimization: whether the test gives models credit for the right things. Giovanni and his co-authors removed near-duplicate questions, rebalanced topic coverage and checked conflicting sources to produce a 1,000-question revision. A wrong reference answer can mark a correct response as a failure. Near-duplicates can give one narrow area of knowledge more weight in the score than the test designers intended.

After DeepMind, Giovanni worked at Thinking Machines Lab before his Cognition affiliation at Code 2026. In his writing on benchmarking, he distinguishes coverage and difficulty. A test full of complex refactors might separate the strongest coding agents. It can still tell a team little about its own workload if users mostly ask for small bug fixes or explanations of existing code. Giovanni recommends reading the tasks themselves before deciding what the score tells you.