Anca Dragan

Training LLM Agents to Empower Humans

CTRL-Rec: Controlling Recommender Systems With Natural Language

Benchmarking Anomaly Detection for Large Language Model Alignment

Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following

Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking

Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning

Context Steering: Controllable Personalization at Inference Time

On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback

AssistanceZero: Scalably Solving Assistance Games

Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL

Adversaries Can Misuse Combinations of Safe Models

Interactive Dialogue Agents via Reinforcement Learning with Hindsight Regenerations

Successor Representations Enable Emergent Compositional Instruction Following

Adversaries Can Misuse Combinations of Safe Models

Reliability-Aware Preference Learning for LLM Reward Models

Defining Deception in Decision Making

Zero-Shot Goal Dialogue via Reinforcement Learning on Imagined Conversations