Dylan Hadfield-Menell
Associate Professor
Studies AI alignment, including preference learning, the limitations of reinforcement learning from human feedback, and evaluation of cultural alignment.
Dylan Hadfield-Menell's lab at MIT works on trustworthy & safe ML, reinforcement learning and interpretability, evaluation & safety of LMs. 27 papers since 2023, 6 at top venues, mostly NeurIPS and ICLR. 5 PhD students and 1 master's student. 21 alumni, most often to Google and the White House Office of Science and Technology Policy. Applications close .
At a glance
- papers, 2023–2026
- 7/ year
- at T1 venues
- 22%
- PhD students
- 5
- MS students
- 1
- alumni listed
- 21
Papers by yearfrom the lab homepage and DBLP, 2023–2026
- 5 papers2026 · T1 1 · collecting
- 7 papers2025 · T1 2
- 8 papers2024 · T1 2
- 7 papers2023 · T1 1
T1 venuesOtherVenues: NeurIPS, ICLR, AAAI/ACM AIES, ICML
Research focuspaper titles that use each topic's words
- Reinforcement learning7 papers
- Interpretability, evaluation & safety of LMs2 papers
- alignment2 papers
- preference2 papers
- Trustworthy & safe ML0 papers
Paper list
From the lab homepage and DBLP.
20265 papers
- —Activation Steering via Generative Causal Mediation (opens in a new tab)
- —Building AI for the Democratic Matrix: A Technical Research Agenda for Normative Competence and Normative Institutions (opens in a new tab)
- Trans. Mach. Learn. Res.Open Technical Problems in Open-Weight AI Model Risk Management (opens in a new tab)
- ICML (T1)Prompt Injection as Role Confusion (opens in a new tab)
- AAAI/ACM AIESSafety Drift After Fine-Tuning: Evidence from High-Stakes Domains (opens in a new tab)
20257 papers
- Trans. Mach. Learn. Res.Defending Against Unforeseen Failure Modes with Latent Adversarial Training (opens in a new tab)
- ICLR (T1)Diverse Preference Learning for Capabilities and Alignment (opens in a new tab)
- NeurIPS (T1)Evaluating Generalization Capabilities of LLM-Based Agents in Mixed-Motive Scenarios Using Concordia (opens in a new tab)
- Trans. Mach. Learn. Res.Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs (opens in a new tab)
- Trans. Mach. Learn. Res.Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities (opens in a new tab)
- —Pitfalls of Evidence-Based AI Policy (opens in a new tab)
- Proceedings of theRandomness, Not Representation: The Unreliability of Evaluating Cultural Alignment in LLMs (opens in a new tab)
20248 papers
- Proceedings of theBlack-Box Access is Insufficient for Rigorous AI Audits (opens in a new tab)
- Trans. Recomm. Syst.Building Human Values into Recommender Systems: An Interdisciplinary Synthesis (opens in a new tab)
- ICLR (T1)Distributional Preference Learning: Understanding and Accounting for Hidden Context in RLHF (opens in a new tab)
- Auton. Agents Multi Agent Syst.Formal Contracts Mitigate Social Dilemmas in Multi-Agent Reinforcement Learning (opens in a new tab)
- NeurIPS (T1)Melting Pot Contest: Charting the Future of Generalized Cooperative Intelligence (opens in a new tab)
- —Red Teaming AI: The Devil Is In The Details (opens in a new tab)
- —The Concordia Contest: Advancing the Cooperative Intelligence of Language Agents (opens in a new tab)
- Proceedings of theThe SaTML 24 CNN Interpretability Competition: New Innovations for Concept-Level Interpretability (opens in a new tab)
20237 papers
- Proceedings of theCognitive Dissonance: Why Do Language Model Outputs Disagree with Internal Representations of Truthfulness? (opens in a new tab)
- Proceedings of theGet It in Writing: Formal Contracts Mitigate Social Dilemmas in Multi-Agent RL (opens in a new tab)
- —Judging Facts, Judging Norms: Training Machine Learning Models to Judge Humans Requires a Modified Approach to Labeling Data (opens in a new tab)
- Trans. Mach. Learn. Res.Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback (opens in a new tab)
- NeurIPS (T1)Red Teaming Deep Neural Networks with Feature Synthesis Tools (opens in a new tab)
- SaTMLToward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networks (opens in a new tab)
- SafeAI@AAAIWhite-Box Adversarial Policies in Deep Reinforcement Learning (opens in a new tab)
Where alumni went21 listed, 15% to industry
Industry 2Academia 11
Most often to
- the White House Office of Science and Technology Policy
Also to
- MIT
- Harvard University
- Harvey Mudd College. Now: a Ph.D. student in computer science at Brown University
Apply to MITFall 2027
MIT EECS Computer Science PhD
- Round
- Single round
- Opens
- —
- GRE
- See program page
- English
- TOEFL/IELTS
- Fee
- USD 90
Confirmed on the program page.
Similar labsnearest by research summary
- GBImperial
Matthew Wicker
- Trustworthy & safe ML
- CNZhejiang
Zhikun Zhang
- Trustworthy & safe ML
- HKCUHK
Songtao Lu
- Learning theory & optimization
- CNSJTU
Yang Yang
- Learning theory & optimization
- USUT Austin
Inderjit Dhillon
- Recognition, detection & segmentation
- GBImperial
Eleonora Giunchiglia
- Reasoning & neuro-symbolic AI
Your notes
Save this lab to keep notes.
Data
Checked against the lab homepage.
- Collected
- Verified
- Papers
- The lab homepage and DBLP
- Members
- Lab homepage
- Alumni
- Lab homepage
Tiers are venue classifications, not lab rankings.
MITcloses