(*) Equal contribution

AI Safety

Most modern LLMs are trained to maximize human feedback. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies.

On Targeted Manipulation and Deception when Optimizing LLMs for User Feedback Marcus Williams, Micah Carroll, Adhyyan Narang, Constantin Weisser, Brendan Murphy, Anca Dragan. ICLR 2024 (arxiv)

Sequential and Strategic Learning

Traditional ML usually considers the learner to be an independent agent in an isolated world; and chooses training objectives and algorithms accordingly. However, in practice data is generated by humans, and models are deployed in complex ecosystems with many other models. The decisions of any one learner has an impact on other learners and users. How can we ensure that these systems exhibit long-run behavior that is socially desirable?

Sample Complexity Reduction via Policy Difference Estimation in Tabular Reinforcement Learning Adhyyan Narang, Andrew Wagenmaker, Lillian J. Ratliff, Kevin Jamieson. NeurIPS 2024 (arxiv)

Online SuBmodular + SuPermodular (BP) Maximization with Bandit Feedback Adhyyan Narang, Omid Sadeghi, Lillian J Ratliff, Maryam Fazel, Jeff Bilmes. UAI 2024 (arxiv)

Decision Dependent Learning in the Presence of Competition Adhyyan Narang, Evan Faulkner, Dmitriy Drusvyatskiy, Maryam Fazel, Lillian J Ratliff. AIStats 2022 (arxiv)

Global Convergence to Local Minmax Equilibrium in Classes of Nonconvex Zero-Sum Games. Tanner Fiez, Lillian J. Ratliff, Eric Mazumdar, Evan Faulkner, Adhyyan Narang. Neurips 2021 (NeurIPS)

Overparameterized Learning

In machine learning, the loss on training data is often used as a proxy for the loss on an unseen test point. Traditional statistical theory justifies this approach for underparameterized settings: when the number of training points greatly exceeds the number of parameters of the model. However, modern neural networks are almost always overparameterized. Is the above proxy still a reasonable choice?

We explore this question for classification problems. We show that classification is ‘easier’ than regression, and that overparameterization could make models brittle even when regular performance is good. Additionally, we study meta-learning in overparameterization; this also gives an insight into where the prior for single-agent overparameterized regression comes from.

Classification and Adversarial examples in an Overparameterized Linear Model: A Signal Processing Perspective Adhyyan Narang, Vidya Muthukumar, Anant Sahai. Short version in ICML OPPO Workshop 2021 (arxiv)

Towards Sample-Efficient Overparameterized Meta-Learning Yue Sun, Adhyyan Narang, Ibrahim Gulluk, Samet Oymak, Maryam Fazel. Neurips 2021 (arxiv)

Classification vs regression in overparameterized regimes: Does the loss function matter? Vidya Muthukumar*, Adhyyan Narang*, Vignesh Subramanian*, Mikhail Belkin, Daniel Hsu, Anant Sahai. JMLR 2021 (arxiv)