A undergraduate student in Testing, Database, Graph.
Agents can spend test-time compute by trying, observing, and revising. We derive an Elo reference for repeated sampling, then show that in a 2022 two-week coding marathon, current agents plateau within 24 hours while top humans keep improving.
AI benchmarks decide which models get funded, deployed, and trusted. We hacked 13 of them. 45 working exploits. Every benchmark rated critical. If the scores are fake, so is everything built on them — including your training data.
We present Argus, a novel framework that uses LLMs to automatically discover test oracles for DBMS testing — finding 41 previously unknown bugs across 5 widely-used databases. Accepted at SIGMOD 2026.