Pierluca D'Oro
Bio
Pierluca D’Oro is founder at a stealth startup revolutionizing how humans interact with AI-generated software. At Mila, I pioneered two early ideas that now sit at the center of agent development: making reinforcement learning scale through simple recipes, and using LLMs as feedback systems to train agents. At Meta Superintelligence Labs, I worked on frontier model development and led environment generation for mobile computer use agents.
Session (1)
Pierluca D'Oro is an AI researcher (most recently at Meta FAIR, after a Mila PhD) whose recent work builds the scientific foundations for human-centered AI agents — from reward-guided decoding for multimodal LLMs (ICCV 2025) to a capability framework for agent design (ADEPTS, 2025) and general-purpose model-free RL (2025). His session is worth attending for a principled, research-backed perspective on what it takes to make AI agents trustworthy and controllable in practice.
Recent talks (1)
GitHub
@proceduralia- proceduralia/pytorch-GAN-timeseries★ 274GANs for time series generation in PyTorch
- proceduralia/pytorch-neural-enhance★ 47Experiments on CNN-based image enhancement in PyTorch
- proceduralia/high_replay_ratio_continuous_control★ 28Efficient seed-parallel implementation of 'Breaking the Replay Ratio Barrier'
- facebookresearch/motiforg★ 135Intrinsic Motivation from Artificial Intelligence Feedback
Recent writing (6)
- ADEPTS: A Capability Framework for Human-Centered Agent Design · paper · Jul 2025
- DigiData: Training and Evaluating General-Purpose Mobile Control Agents · paper · Nov 2025
- Towards General-Purpose Model-Free Reinforcement Learning · paper · Jan 2025
- Mol-MoE: Training Preference-Guided Routers for Molecule Generation · paper · Feb 2025
- Controlling Multimodal LLMs via Reward-guided Decoding · paper · Aug 2025
- Science of AI Agents (Substack) · blog
Public activity researched automatically · as of Jun 2026