MIT Algorithmic Alignment Group

The Algorithmic Alignment Group is a research program at MIT CSAIL (Computer Science and Artificial Intelligence Laboratory), established when Dylan Hadfield-Menell joined MIT as an Assistant Professor in 2021. Hadfield-Menell, who holds the Bonnie and Marty Tenenbaum Career Development Professorship and was promoted to Associate Professor in 2024, directs the group within CSAIL's Embodied Intelligence cluster. The group's mission is to develop better conceptual understanding, algorithmic techniques, and policies to make AI more safe and socially beneficial. Their work is interdisciplinary and centers on how humans and AI systems interact across the contexts of value learning, incentives, recommendation systems, debugging, and policy. Core research areas include: AI safety and robustness (adversarial attacks, unlearning methods, latent adversarial training); AI alignment and preference learning (reinforcement learning from human feedback, assistance games, goal inference); interpretability and transparency (neural network diagnostics, feature analysis, internal representations of language models); responsible AI development (auditing, open-weight model risk, deepfake harms); and AI governance and policy (evidence-based policy design, recommendation system regulation, multi-agent coordination). Notable outputs include the CommonClaim dataset of 20,000 human-labeled statements for red-teaming language models, a contracts library for multi-agent reinforcement learning, and influential publications on cooperative inverse reinforcement learning, latent adversarial training, and the limitations of reinforcement learning from human feedback. The group currently includes a postdoctoral researcher, four PhD students, and over a dozen Masters/M.Eng. students. Dylan Hadfield-Menell was named an AI2050 Early Career Fellow by Schmidt Sciences in 2022, supporting his work on AI systems that manage uncertainty about rewards.