For a full list, see Google Scholar.
Evaluation is an inference problem: observed behavior is evidence about capabilities and harms that cannot be measured directly. We develop metrics, scoring rules, and protocols that specify what is being measured, establish what the evidence can support about it, and tie those conclusions to the decisions and societal consequences that ultimately matter.
Optimization and sampling determine how models are trained and how uncertainty is quantified at scale. We study the geometric and dynamical structure these problems share, using it to design faster algorithms and to preserve accuracy when learning must operate under real constraints, such as privacy, limited computation, and reduced precision.
AI systems increasingly mediate what people create, believe, and receive. We develop technical methods for detecting and controlling harmful model behavior, and mathematical, empirical, and normative frameworks for understanding how these systems shape human agency, epistemic diversity, and access to opportunity.
Constructs and valid operationalizations.
Benchmark scores are often treated as direct measurements of what an AI system can do, but interpreting performance as evidence of capability requires assumptions about how an underlying capability produces observed behavior. We argue that evaluations should make this inferential step explicit by beginning with a theory of capability and developing methods that quantify uncertainty about the capability being measured.
Benchmark scores are often treated as direct measurements of what an AI system can do, but interpreting performance as evidence of capability requires assumptions about how a latent capability produces observed behavior. We argue that evaluations should make this inferential step explicit by beginning with a theory of capability and developing methods that quantify uncertainty about the capability being measured.
AI assistants are typically evaluated by how well they satisfy a user’s request, but users often approach complex tasks before their goals and intentions are fully formed. We argue that effective assistance should instead account for the user’s evolving cognitive process, motivating evaluations that ask not only whether a task was completed, but whether the system preserved appropriate human judgment and agency throughout the interaction.
AI assistants are typically evaluated by whether they satisfy a user’s request, but this operationalization assumes that users enter complex tasks with fixed and fully formed goals. We argue that successful assistance should instead be defined in relation to the user’s evolving cognitive process, including whether the system preserves appropriate human judgment and agency.
Context dependence, robustness, and the limits of inference.
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.
Some harmful model capabilities cannot be safely or lawfully evaluated by generating outputs, making the usual behavioral evidence unavailable. We develop Evaluation without Generation, the first machine-learning screening method evaluated on CSAM-specialized LoRAs found in the wild, detecting harmful model specialization without generating or inspecting CSAM.
Shapley values and other semivalues are often used to rank training examples for data selection, but these rankings can depend on arbitrary choices in how model utility is specified. We characterize when the resulting selection decisions are invariant to this utility ambiguity, and develop algorithms that either certify robustness or produce explicit witnesses showing that the selected data are not uniquely determined by the underlying task.
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.
Context-free evaluations treat behavior observed on isolated prompts as evidence of how a system will behave during interaction, but information about the user and prior exchanges can change the behavior being measured. We show that interaction context often increases sycophancy, demonstrating that context-free protocols can systematically understate risks to truthfulness and user autonomy.
Machine unlearning methods are typically evaluated immediately after a concept has been removed, implicitly treating successful forgetting as a static property of the model. We show that subsequent benign fine-tuning can cause supposedly unlearned concepts to reappear, revealing that evaluations of unlearning must test whether safety interventions remain effective under realistic sequences of model updates.
Unlearning is commonly evaluated immediately after an intervention, implicitly treating successful suppression as evidence of durable forgetting. We show that subsequent benign fine-tuning can cause unlearned concepts to reappear for diffusion models, demonstrating that evaluations must test whether safety properties persist through realistic sequences of model updates.
Influence functions estimate how perturbations to training data affect a learned model, but widely used Fisher-information-based approximations have lacked a clear account of when they are accurate. We characterize the asymptotic accuracy of Fisher-based influence estimators, compare them with Hessian-based methods, and extend the analysis to learning problems with nondifferentiable regularization.
Decision-relevant metrics and downstream consequences.
Classification models are often compared using fixed thresholds or aggregate metrics that can be poorly matched to how their predictions will actually be used. We develop a decision-theoretic framework for choosing evaluation metrics based on the downstream consequences of prediction, showing how properties of the deployment setting determine which measures meaningfully distinguish between models.
Classification models are often compared using accuracy, AUC, or fixed thresholds, even when those metrics are poorly matched to how predictions will actually be used. We develop a decision-theoretic framework that derives evaluation metrics from downstream consequences, showing how the deployment setting determines which distinctions between models matter.
Standard classification metrics evaluate predictions independently, even when predictions are used to determine the order in which alternatives are searched or tested. We develop Pandora’s Regret, a proper scoring rule derived from the downstream cost of sequential search, aligning model evaluation with the decisions that predictions ultimately support.
Standard scoring rules evaluate predictions independently, but predictions used for sequential search matter through the order and cost of the resulting search process. We develop Pandora’s Regret, a proper scoring rule derived from downstream search costs, aligning model evaluation with the decision problem that predictions support.
Clinical prediction systems operate under changing disease prevalence and asymmetric costs of different errors, yet common metrics such as accuracy and AUC can ignore these deployment realities. We develop a decision-theoretic evaluation framework that incorporates calibration, label shift, and clinically meaningful error costs, enabling model comparisons that better reflect the conditions and priorities of real clinical use.
Accuracy and AUC can obscure calibration, label shift, and asymmetric error costs that determine clinical value. Building on the theory of proper scoring rules, we derive new bounded, cost-sensitive log scores that connect calibration to decision performance across clinically relevant ranges of class prevalence, providing principled metrics for comparing models under uncertain deployment conditions.
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.
Equal opportunity cannot always be evaluated from individual decisions in isolation because multiple systems collectively determine which pathways to opportunity remain available. We develop algorithmic pluralism as a framework for evaluating the diversity and legitimacy of these pathways, shifting attention from isolated predictions to the structural bottlenecks created by an algorithmic ecosystem.
When many models perform equally well, model multiplicity is often taken to imply greater flexibility, fairness, or diversity in downstream decisions. We show that multiplicity in predictive models need not translate into multiplicity in the allocations that ultimately matter, motivating evaluation at the level of downstream opportunities rather than model predictions alone.
Model multiplicity is often used as evidence that decision-makers have meaningful flexibility among equally accurate models, but differences between models need not produce differences in allocations. We show that evaluation must examine the opportunities people actually receive, distinguishing predictive multiplicity from allocation multiplicity.
Data-valuation methods are often evaluated through aggregate model performance, even though using those values to curate training data can redistribute errors and benefits across groups. We study how semivalue-based curation affects predictive performance and fairness, evaluating data-selection methods through the downstream outcomes they produce.
Data-valuation methods are often evaluated through aggregate model performance, even though using those values to curate training data can redistribute errors and benefits across groups. We study how semivalue-based curation affects predictive performance and fairness, evaluating data-selection methods through the downstream outcomes they produce.
Geometry and dynamics for optimization and sampling.
Optimization and sampling are closely connected through the continuous-time dynamics that underlie their algorithms. We show that Hamiltonian dynamics, a central tool in modern sampling, can also serve as a primitive for deterministic accelerated optimization by exploiting contraction of averaged trajectories, yielding practical algorithms with optimal first-order complexity.
The performance of optimization algorithms often depends critically on choosing appropriate step sizes, yet standard backtracking adjusts them using a fixed rule that ignores how badly a candidate step fails. We develop an adaptive line search that uses the magnitude of this violation to choose better steps, improving practical optimization speed without sacrificing the convergence guarantees of classical methods.
Many inference problems require sampling from high-dimensional distributions subject to complex constraints, where standard Euclidean sampling methods can be inefficient. We develop a geometry-aware Langevin sampler inspired by preconditioned and natural-gradient optimization, showing how optimization geometry can be transferred to sampling to obtain high-accuracy guarantees for distributions supported on constrained spaces.
Sampling from distributions constrained to compact convex sets is challenging because standard Euclidean methods can handle boundary geometry poorly, while existing Mirror Langevin discretizations introduce asymptotic bias. We develop the Metropolis-adjusted Mirror Langevin algorithm, which combines mirror geometry with an accept-reject correction to obtain unbiased sampling and mixing guarantees with exponentially better dependence on the target accuracy.
Mean-field optimization provides a useful formulation for problems such as training mean-field neural networks, but turning continuous-time dynamics into practical algorithms requires both particle and time discretization. We develop an N-particle underdamped Langevin algorithm, establish fast mixing for the underlying dynamics, and prove global convergence of its spacetime discretization in total variation.
Many nonconvex learning problems have additional structure that makes global optimization possible even without convexity. We develop accelerated stochastic methods for smooth quasar-convex objectives, providing improved convergence guarantees and a unified analysis that also covers existing deterministic methods, with applications including learning linear dynamical systems.
Preserving utility under privacy and resource constraints.
AI agents increasingly rely on prompts, tools, and workflows that can themselves be improved over time, but they must learn from imperfect feedback: observing a failure does not reveal which proposed change will actually help. We show that simply revising these components based on past failures can reinforce incorrect lessons. VERSE uses execution-based verification to turn uncertain feedback into reliable evidence before incorporating changes, leading to more reliable gains across several agent-optimization methods.
Modern language models can be expensive to store and deploy, and standard quantization methods often leave the large softmax output layer in high precision. We formulate softmax quantization as a class-aware rate-distortion problem and develop SoftWater, which allocates precision according to how different classes affect the model’s output distribution, substantially reducing memory while preserving predictive quality.
Randomized sketching can make large optimization problems cheaper to solve while also providing privacy through its inherent randomness, but existing private sketches can remain computationally expensive. We develop a privacy mechanism based on fast randomized transforms that retains strong privacy and accuracy guarantees while enabling faster private linear regression, connecting computational efficiency and privacy within a common algorithmic primitive.
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.
Randomization used to make optimization more computationally efficient can also provide formal privacy guarantees. We develop a sharper privacy analysis of Gaussian sketching and show how the same randomized compression operation can simultaneously support efficient computation and differential privacy, leading to improved accuracy for private linear regression and, in some settings, reduced runtime.
Machine unlearning aims to remove the effect of deleted training examples without retraining a model from scratch, but existing approximate methods can be computationally costly and often assume smooth learning objectives. We develop an efficient online unlearning method based on the infinitesimal jackknife that extends to nonsmooth regularizers, and show how standard model-selection procedures such as cross-validation can invalidate existing approximate unlearning guarantees.
Non-generative auditing, model steering, and durable unlearning.
Some harmful model capabilities cannot be safely or lawfully evaluated by generating outputs, making the usual behavioral evidence unavailable. We develop Evaluation without Generation, the first machine-learning screening method evaluated on CSAM-specialized LoRAs found in the wild, detecting harmful model specialization without generating or inspecting CSAM.
Activation steering offers an efficient way to control LLM behavior without retraining, but existing methods are brittle because concept-related features can vary across tokens, activations, and layers. We develop an attention-guided framework that automatically identifies the features and layers most relevant for steering, nearly doubling the number of successfully steered concepts across model architectures and sizes up to 70 billion parameters.
Machine unlearning methods are typically evaluated immediately after a concept has been removed, implicitly treating successful forgetting as a static property of the model. We show that subsequent benign fine-tuning can cause supposedly unlearned concepts to reappear, revealing that evaluations of unlearning must test whether safety interventions remain effective under realistic sequences of model updates.
Removing sensitive or undesirable information from large language models can require expensive retraining and can degrade performance on unrelated tasks. We develop UCD, an inference-time unlearning method that uses contrastive decoding between smaller auxiliary models to suppress targeted information while preserving the original model’s utility, improving both forgetting and retained performance on standard unlearning benchmarks.
Sycophancy, co-creation, and epistemic diversity.
AI assistants are typically evaluated by how well they satisfy a user’s request, but users often approach complex tasks before their goals and intentions are fully formed. We argue that effective assistance should instead account for the user’s evolving cognitive process, motivating evaluations that ask not only whether a task was completed, but whether the system preserved appropriate human judgment and agency throughout the interaction.
One-shot image generation can make visual decisions before a user’s intentions are fully formed, prematurely anchoring the creative process and reducing the user’s control over subsequent edits. We develop Creo, a multi-stage system that progresses from editable sketches to high-fidelity images while preserving prior decisions, enabling greater exploration, stronger user ownership, and less homogeneous outputs than conventional one-shot generation.
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.
Whether homogeneous language-model outputs are harmful depends on the task: creative writing may benefit from variation in narrative content, while mathematical reasoning may require agreement on a verifiable answer. We develop and validate a taxonomy of task-dependent functional diversity, then introduce a sampling method that increases diversity where it is valuable without imposing a task-agnostic tradeoff between diversity and quality.
Pluralism, competition, and consequential allocation.
Data-valuation methods are often evaluated through aggregate model performance, even though using those values to curate training data can redistribute errors and benefits across groups. We study how semivalue-based curation affects predictive performance and fairness, evaluating data-selection methods through the downstream outcomes they produce.
Firms increasingly rely on similar data, objectives, and pretrained models, causing their pricing algorithms to produce correlated predictions even without direct coordination. We model this phenomenon in personalized pricing and show that greater algorithmic correlation reduces consumer welfare and can incentivize firms to sacrifice predictive accuracy for implicit coordination as consumers become more price sensitive.
When many models perform equally well, model multiplicity is often taken to imply greater flexibility, fairness, or diversity in downstream decisions. We show that multiplicity in predictive models need not translate into multiplicity in the allocations that ultimately matter, motivating evaluation at the level of downstream opportunities rather than model predictions alone.
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.
Machine-learning systems used to allocate scarce resources often convert uncertain predictions into deterministic decisions, denying individuals with legitimate competing claims any meaningful chance of receiving an opportunity. We argue that fair allocation can require randomization and develop stochastic procedures that better account for the claims individuals have to scarce social goods and opportunities.
One-shot image generation can make visual decisions before a user’s intentions are fully formed, prematurely anchoring the creative process and reducing the user’s control over subsequent edits. We develop Creo, a multi-stage system that progresses from editable sketches to high-fidelity images while preserving prior decisions, enabling greater exploration, stronger user ownership, and less homogeneous outputs than conventional one-shot generation.
Activation steering offers an efficient way to control LLM behavior without retraining, but existing methods are brittle because concept-related features can vary across tokens, activations, and layers. We develop an attention-guided framework that automatically identifies the features and layers most relevant for steering, nearly doubling the number of successfully steered concepts across model architectures and sizes up to 70 billion parameters.
Firms increasingly rely on similar data, objectives, and pretrained models, causing their pricing algorithms to produce correlated predictions even without direct coordination. We model this phenomenon in personalized pricing and show that greater algorithmic correlation reduces consumer welfare and can incentivize firms to sacrifice predictive accuracy for implicit coordination as consumers become more price sensitive.
Machine-learning systems used to allocate scarce resources often convert uncertain predictions into deterministic decisions, denying individuals with legitimate competing claims any meaningful chance of receiving an opportunity. We argue that fair allocation can require randomization and develop stochastic procedures that better account for the claims individuals have to scarce social goods and opportunities.
Whether homogeneous language-model outputs are harmful depends on the task: creative writing may benefit from variation in narrative content, while mathematical reasoning may require agreement on a verifiable answer. We develop and validate a taxonomy of task-dependent functional diversity, then introduce a sampling method that increases diversity where it is valuable without imposing a task-agnostic tradeoff between diversity and quality.
Removing sensitive or undesirable information from large language models can require expensive retraining and can degrade performance on unrelated tasks. We develop UCD, an inference-time unlearning method that uses contrastive decoding between smaller auxiliary models to suppress targeted information while preserving the original model’s utility, improving both forgetting and retained performance on standard unlearning benchmarks.