Pierluca D’Oro is Programma Labs’ founder and an AI researcher focused on agents, reinforcement learning and evaluation. His recent work examines whether computer-use benchmarks measure adaptation or reward replaying a successful trace.[1][5]
Start with Computer Use at the Edge of the Statistical Precipice →
Education
Links
Website CV
GitHub LinkedIn Google Scholar Writing 
X · proceduralia Founder · Programma Labs
Career
D’Oro trained in engineering in Catania and Milan, then completed doctoral research in Montréal. His research internships and Meta appointments span reinforcement learning, model-based learning and computer-use agents. Programma Labs now focuses on making AI-generated software verifiable by people and machines.[1][2][3]
Contributions
The computer-use paper records a successful trajectory and replays its clicks and typing without observing the screen. On deterministic tasks this can match or exceed the source model. The result challenges the interpretation of pass@k: finding one successful attempt can also supply a reusable tape. The authors propose varied but integrity-checked environments, privileged verification, realistic behavior and isolation. They also use uncertainty estimates that account for task and configuration variation rather than treating repeated runs as the only source of uncertainty.[5]
Tools, writing & talks
Writing, research & talks
Start here
The paper behind blind-replay comparisons and proposals for more reliable computer-use evaluation.
Start with the blind-replay construction and environment-design principles. Basic agent evaluation and probability help; read the stated determinism conditions before applying the pass@k comparison elsewhere.
Coauthor[5]A sub-megabyte script that never looks at the screen, built by replaying one recorded trajectory per task, tops deterministic benchmarks like OSWorld, and the paper proves pass@k on a deterministic environment is exactly that replay's success rate.
Catalog workOriginal sourceA presentation explaining replay exploits, environment variation and statistical uncertainty.
Speaker[6]