Publications

SoftWater: Class-Aware Rate Allocation for Softmax Quantization

with Joao Cavalcanti. arXiv · ∇ Optimization

The Fast Mixing Mechanism for Differential Privacy

with Omri Lev, Moshe Shenfeld, Vishwak Srinivasan, Katrina Ligett. arXiv · ∇ Optimization · ◆ Safety

Alignment has a Fantasia Problem

with Nathanael Jo, Zoe De Simone, Mitchell Gordon. arXiv · ✓ Evaluation

Pandora's Regret: A Proper Scoring Rule for Evaluating Sequential Search

with Gerardo Flores, Yash Deshpande, Jannis R. Brea. arXiv · ✓ Evaluation

Creo: From One-Shot Image Generation to Progressive, Co-Creative Ideation

with Zoe De Simone, Angie Boggust, Fredo Durand, Arvind Satyanarayan. arXiv · ◆ Safety

Evaluation without Generation: Non-Generative Assessment of Harmful Model Specialization with Applications to CSAM

with Vinith Suriyakumar, Ayush Sekhari, Lena Stempfle, Robertson Wang, Michael Simpson, Rebecca Portnoff, Marzyeh Ghassemi. arXiv · ◆ Safety · ✓ Evaluation

Some harmful model capabilities are difficult or unsafe to evaluate by generating outputs, especially when producing the content itself is legally or ethically constrained, as with child sexual abuse material (CSAM). We develop methods for detecting harmful model specialization from a model's internal representations, enabling evaluation without generation.

Accelerated Convex Optimization via Hamiltonian Dynamics with Deterministic Integration Time

with Xiuyuan Wang, Vishwak Srinivasan, Qiang Fu, Siddharth Mitra, Andre Wibisono. arXiv · ∇ Optimization

A Consequentialist Critique of Binary Classification Evaluation: Theory, Practice, and Tools

with Gerardo Flores, Abby Schiff, Alyssa Smith, Julia Fukuyama. arXiv · ✓ Evaluation

Interaction Context Often Increases Sycophancy in LLMs

with Shomik Jain, Charlotte Park, Matheus Vianna, Dana Calacci. arXiv · ◆ Safety · ✓ Evaluation

AI systems increasingly personalize their responses using information about users and their prior interactions, but this context can also amplify problematic forms of agreement and mirroring. We show that interaction context often increases sycophancy, highlighting how context-free evaluations can underestimate behaviors that raise concerns about user autonomy and psychological harm in real-world interactions.

Position: AI Evaluations Should be Grounded on a Theory of Capability

with Nathan Jo. arXiv · ✓ Evaluation

Near-Optimal Private Linear Regression via Iterative Hessian Mixing

with Omri Lev, Vishwak Srinivasan, Moshe Shenfeld, Katrina Ligett. arXiv · ∇ Optimization · ◆ Safety

Differential privacy makes it possible to learn from sensitive data while providing formal guarantees about individual privacy, but these guarantees can come at a substantial cost to statistical accuracy. We develop Iterative Hessian Mixing, a Gaussian-sketching approach to private linear regression that achieves near-optimal accuracy and improves over state-of-the-art methods.

Efficient and accurate steering of Large Language Models through attention-guided feature learning

with Parmida Davarmanesh, Adityanarayanan Radhakrishnan. arXiv · ◆ Safety

Impacts of Semivalue-Based Data Curation on Fair Machine Learning Outcomes

with Hannah Diehl, Justin Steil. paper · ✓ Evaluation · ◆ Safety

Aligning Evaluation with Clinical Priorities: Calibration, Label Shift, and Error Costs

with Gerardo Flores, Alyssa Smith, Julia Fukuyama. arXiv · ✓ Evaluation

The Gaussian Mixing Mechanism: Renyi Differential Privacy via Gaussian Sketches

with Omri Lev, Vishwak Srinivasan, Moshe Shenfeld, Katrina Ligett, Ayush Sekhari. arXiv · ◆ Safety · ∇ Optimization

Homogeneous Algorithms Can Reduce Competition in Personalized Pricing

with Nathan Jo, Kathleen Creel, Manish Raghavan. arXiv · ◆ Safety

Adaptive Backtracking Line Search

with Laurent Lessard, Joao Cavalcanti. arXiv · ∇ Optimization

High-accuracy sampling from constrained spaces with the Metropolis-adjusted Preconditioned Langevin Algorithm

with Vishwak Srinivasan, Andre Wibisono. arXiv · ∇ Optimization

Allocation Multiplicity: Evaluating the Promises of the Rashomon Set

with Shomik Jain, Margaret Wang, Kathleen Creel. arXiv · ◆ Safety · ✓ Evaluation

Algorithmic Pluralism: A Structural Approach To Equal Opportunity

with Shomik Jain, Vinith Suriyakumar, Kathleen Creel. arXiv · ◆ Safety

What equal opportunity requires cannot always be determined by evaluating individual algorithmic decisions in isolation; it depends on how systems collectively structure access to opportunity. We develop algorithmic pluralism as a normative and evaluative framework, arguing that systems should preserve multiple viable pathways to opportunity and be evaluated by the severity and legitimacy of the bottlenecks they create.

Scarce Resource Allocations That Rely On Machine Learning Should Be Randomized

with Shomik Jain, Kathleen Creel. arXiv · ◆ Safety

Unstable Unlearning: The Hidden Risk of Concept Resurgence in Diffusion Models

with Vinith Suriyakumar, Rohan Alur, Ayush Sekhari, Manish Raghavan. arXiv · ◆ Safety · ✓ Evaluation

LLM Output Homogenization is Task Dependent

with Shomik Jain, Jack Lanchantin, Maximilian Nickel, Karen Ullrich, Jamelle Watson-Daniels. arXiv · ◆ Safety

UCD: Unlearning in LLMs via Contrastive Decoding

with Vinith M. Suriyakumar, Ayush Sekhari. arXiv · ◆ Safety

Adaptive kernel selection for Stein Variational Gradient Descent

with Moritz Melcher, Simon Weissmann, Jakob Zech. arXiv · ∇ Optimization