Pierluca D'Oro

Pierluca D’Oro is the founder of Programma Labs, which develops infrastructure for scalable human oversight of AI-generated software. His research spans reinforcement learning, reward design, and the evaluation of autonomous computer-use agents, with particular attention to systems that appear capable because their training signals or benchmarks are misleading.

D’Oro studied engineering in Italy and completed his PhD at Mila in Montreal. During his doctoral work, he joined Meta, where he spent more than three years and contributed to research at Meta Superintelligence Labs before founding Programma Labs. His early projects included generative modeling for time-series data and neural image enhancement; his research increasingly concentrated on helping agents learn effectively from imperfect experience.

His work on primacy bias in deep reinforcement learning showed how agents can overfit their earliest interactions and resist better evidence acquired later. Periodically resetting selected model components helped counteract that lock-in. The same insight informed research on breaking the replay ratio barrier: improving sample efficiency by learning more aggressively from previously collected experience without destabilizing training. D’Oro published an implementation for high-replay-ratio continuous control alongside that work.

With collaborators, he developed Motif, which converts language-model preferences over descriptions of behavior into intrinsic rewards for reinforcement-learning agents. Tested in NetHack, the approach uses a language model as a source of feedback while a separate agent learns to act. D’Oro has argued that rewards cannot always be fully trusted—and that learning under precisely those conditions should be central to reinforcement-learning research.

• Blind replay agents and benchmark exploitation. Deterministic computer-use benchmarks can reward scripts that replay successful sequences of taps, typing, and scrolling without understanding their tasks. D’Oro connected this exploit mathematically to pass@k, demonstrating how familiar metrics can mistake memorized trajectories for genuine capability.

• PRISM principles and the DigiWorld benchmark. At Meta Superintelligence Labs, he helped develop realistic, sandboxed, verifiable evaluation environments with systematic variation. DigiWorld spans 15 Android applications, 387 verified scenarios, and 3.2 million verified configurations; its compiler-like infrastructure assembles task templates, mock data, interface states, and verifiers while rejecting invalid combinations.

• Honest uncertainty in agent evaluation. Agent performance varies across both repeated actions and changing environments. In cases D’Oro analyzed, nominal 95 percent confidence intervals that ignored environmental variation achieved only approximately 17–20 percent actual coverage. Accounting for the benchmark’s hierarchy gives deployment teams a more reliable basis for comparing models and deciding when additional evaluation is necessary.

At AI Engineer World’s Fair 2026, D’Oro described Programma’s focus on verification infrastructure for computer-use agents: applying rigorous evaluation and human oversight to increasingly autonomous software systems.