Explore Our Research

2025

Technical Report

Technical Report: Full-Stack Fine-Tuning for the Q Programming Language

A framework for the full stack finetuning of LLMs (adaptive pretraining, supervised fine-tuning and reinforcement learning), applied to the Q programming language.

 

ACL 2025

Spectra 1.1: Scaling Laws and Efficient Inference for Ternary Language Models

Scaling law analysis of TriLMs and introduces Spectra-1.1, an open suite of TriLMs trained on up to 1.2 trillion tokens.

 

AISTATS 2025

Variational Schrodinger Momentum Diffusion

We include momentum acceleration in the simulation-free transport-optimized diffusion models to further enhance generation quality and simplify the denoising process.

 

AISTATS 2025

Optimal Stochastic Trace Estimation in Generative Modeling

We incorporate deterministic computations of major eigenvalues into stochastic trace estimators to further reduce the training variance in generative models and enhance generation quality.

 

AISTATS 2025

Parabolic Continual Learning

We describe a robust algorithm for continual learning that provides generalization and forgetting bounds by analyzing the dynamics of the loss over new task observations through the perspective of a parabolic partial differential equation.

 

AISTATS 2025

Dissecting the Impact of Model Misspecification in Data-Driven Optimization

We dissect the performance comparisons between two data-driven optimization approaches in terms of the amount of model misspecification.

 

ICLR 2025

Elliptic Loss Regularization

We propose a new type of regularization for general function approximators based on enforcing the loss to satisfy an elliptic operator through a computationally scalable scheme, and we then prove that this regularization provides benefits in terms of uncertainty quantification and robustness to distribution shift.

 

ICML 2025 (Spotlight)

Privacy amplification by structured subsampling for deep differentially private time series forecasting

We adapt DP-SGD to timeseries data, as the guarantees for DP-SGD are incompatible with time series specific tasks like forecasting and they rely on unstructured sampling.

 

ICML 2025 (Spotlight)

Sparse-pivot: Dynamic correlation clustering for node insertions

We substantially improve on the approximation guarantee of the dynamic correlation clustering where at each time step a new node arrives.

 

ICML 2025

Breaking the n^1.5 additive error barrier for private and efficient graph sparsification via private expander decomposition

Given an input graph, we output a synthetic private graph that approximates all cuts of the original graph with error lower than n^1.5.

 

ICML 2025 Workshop on Computer Use Agents

Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment

This paper investigates enhancing the reasoning capabilities of LLM agents using Reinforcement Learning, focusing on multi-turn tool-use scenarios which can be modeled as Markov Decision Processes.

 

NeurIPS 2025

Differentially Private Gomory-Hu Trees

We consider the problem of all pairs s-t cuts, and output a synthetic graph in a differentially private way that can answer min s-t cut queries with optimal additive error.

 

NeurIPS 2025

Efficiently Verifiable Proofs of Data Attribution

We design a theoretical interactive prover-verifier protocol that allows resource-constrained parties to efficiently verify data attributions provided by untrusted and computationally powerful parties, achieving Probably-Approximately-Correct (PAC) guarantees.

 

NeurIPS 2025

TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster

A retrieval-augmented generation (RAG) framework for time series forecasting that enhances the generalization and interpretability of Time Series Forecasting Models.

 

NeurIPS 2025

SAS: Simulated Attention Score

SAS projects low-dimensional head representations and query/key embeddings into higher-dimensional spaces to mimic more attention heads and larger per-head hidden sizes without increasing the model’s parameter count.

 

SODA 2025

Average-Case Hardness of Parity Problems: Orthogonal Vectors, k-SUM and More

This work shows that some of the most fundamental problems are still hard to solve when you take the input randomly from a distribution.

 

TMLR 2025

Covariate-dependent Graphical Model Estimation via Neural Networks with Statistical Guarantees

We investigate a neural-network based approach for estimating covariate-dependent graphical models and provide the corresponding theoretical guarantees.

 

TMLR 2025

Reweighting Improves Conditional Risk Bounds

We investigate the risk bounds associated with a two-step ERM procedure.

 

UAI 2025

Conditional Average Treatment Effect Estimation Under Hidden Confounders

We consider a technique that uses limited randomized control trial data to mitigate hidden confounder bias when estimating conditional average treatment effects.

 

UAI 2025

Off-Policy Predictive Control with Causal Sensitivity Analysis

We improve upon model predictive control to provide control policies that tightly bound the worst-case regret.

 

WACV 2025

PivotAlign: Improve Semi-Supervised Learning by Learning Intra-Class Heterogeneity and Aligning with Pivots

We improve representation learning in semi-supervised learning by enforcing data points' alignment with learned pivot points that represent substructure within each data class.

 

2024

NAACL 2024

Task-Agnostic Detector for Insertion-Based Backdoor Attacks

We propose a task-agnostic trojan detection method for NLP methods by investigating the activation pattern.

 

ICML 2024

Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting

We propose a method to align the pre-trained semantic space learned by LLMs with time series embedding space to perform time series forecasting based on learned prompts from the joint space.

 

ICML 2024

Constrained Exploration via Reflected Replica Exchange Stochastic Gradient Langevin Dynamics

We propose reflected replica exchange stochastic gradient Langevin dynamics for constrained non-convex exploration, which improves naive reSGLD.

 

ICML 2024

Pruned pivot: correlation clustering algorithm for dynamic, parallel, and local computation models

We introduce a simple algorithm for correlation clustering that improves state of the art running times in MPC and dynamic settings.

 

ICML 2024

Variational Schrödinger Diffusion Models

This paper pioneers the exploration of the ADAM alternative to SGD, a vital step for more transport-efficient diffusion models.

 

UAI 2024

Distributionally Robust Optimization as a Scalable Framework to Characterize Extreme Value Distributions

We consider robustifying estimates of multivariate extreme value distributions to better hedge against worst case losses.

 

UAI 2024

Base Models for Parabolic Partial Differential Equations

We develop techniques for solving parabolic PDEs with both high accuracy and fast computation speed for potential use in applications such as derivative pricing and optimal control.

 

UAI 2024

On Convergence of Federated Averaging Langevin Dynamics

We propose federated averaging Langevin algorithm (FA-LD) for uncertainty quantification with distributed clients and studied the convergence in convex scenarios.

 

UAI 2024 (Oral)

Reflected Schrödinger Bridge for Constrained Generative Modeling

We introduce the Reflected Schrodinger Bridge algorithm: an entropy-regularized optimal transport approach tailored for generating data within diverse bounded domains.

 

Journal of Computational and Graphical Statistics

Bayesian Federated Learning with Hamiltonian Monte Carlo: Algorithm and Theory

This work introduces a novel and efficient Bayesian federated learning algorithm, namely, the Federated Averaging stochastic Hamiltonian Monte Carlo (FA-HMC), for parameter estimation and uncertainty quantification.

 

Journal of Computational and Graphical Statistics

Structural Discovery With Partial Ordering Information for Time-Dependent Data With Convergence Guarantees

Built upon existing work, this paper proposes an ADMM-based algorithm that handles the estimation of a linear SEM, in the presence of partial ordering information known as apriori.

 

AISTATS 2024

Accelerating Approximate Thompson Sampling With Underdamped Langevin Monte Carlo

We found that approximate Thompson sampling with underdamped Langevin Monte Carlo is more sample efficient.

 

AISTATS 2024

Graph Partitioning with a Move Budget

Approximation algorithms for k-partitioning when there is an initial partitioning of the network and want to achieve a "good" partitioning while moving as few nodes as possible.

 

AISTATS 2024

Neural McKean-Vlasov Processes: Distributional Dependence in Diffusion Processes

We provide a framework for analyzing neural network architectures, such as the transformer, within the context of stochastic processes.

 

AISTATS 2024

Low-rank MDPs with Continuous Action Spaces

We study the problem of extending PAC algorithms for low-rank MDPs to settings with continuous actions and explore multiple concrete approaches for performing this extension. 

 

NeurIPS 2024

Stopping Bayesian Optimization with Probabilistic Regret Bounds

Principled model-based stopping rules for Bayesian optimization.

 

NeurIPS 2024

Efficient and Sharp Off-Policy Evaluation in Robust Markov Decision Processes

We study statistically efficient evaluation of policies under best- and worst-case perturbations to a Markov decision process (MDP) given offline transition observations, which accounts for unmeasured confounding.

 

NeurIPS 2024 Workshop on Table Representation Learning

Recurrent Interpolants for Probabilistic Time Series Prediction

We propose a new approach to multivariate time series forecasting, combining the strengths of sequential models with diffusion probabilistic modeling, based on stochastic interpolants and conditional generation with control features to better capture high-dimensional distributions and cross-feature dependencies.

 

IJCAI 2024, Survey Track

Empowering Time Series Analysis with Large Language Models: A Survey

This survey provides a systematic overview of various methods that utilize pre-trained large language models for time series analysis, discussing challenges, motivations, and future research opportunities.

 

Quantitative Finance 2024

Do price trajectory data increase the efficiency of market impact estimation?

We consider an efficient method for the market impact estimation problem.

 

ICLR 2024

VQ-TR: Vector Quantized Attention for Time Series Forecasting

We augment the attention mechanism by quantizing the query vectors to obtain a novel attention block for forecasting.

 

STOC 2024

Listing Cliques From Smaller Cliques

Explore our study centered on finding an output-sensitive listing of k-cliques in networks. 

 

TMLR 2024

A VAE-based Framework for Learning Multi-Level Neural Granger-Causal Connectivity

We consider the problem of estimating neural Granger causality in the presence of entity-specific heterogeneity.

 

IEEE Transactions on Signal Processing 2024

A Communication-Efficient Algorithm for Federated Multilevel Stochastic Compositional Optimization

We consider the multilevel stochastic composite optimization problem in a distributed setting.

2023

EMNLP 2023

Attention-Enhancing Backdoor Attacks Against BERT-based Models

We propose a plug-and-play loss module that can easily boost a text-based attacking algorithm's performance whenever the victim is any transformer-like architecture with attention modules. 

 

TMLR 2023

Learning To Abstain in the Presence of Uninformative Data

When the majority of the data is highly noisy, we propose a model for selecting data that is predictable, learnable and informative.

 

ICLR 2023 Workshop

On the Existence of a Trojaned Twin Model

We build a theoretic framework for UAP-based backdoor attacks. This paper proposes the concept of a Trojan twin model and a practical heuristic algorithm.

 

AAAI 2023

Non-Reversible Parallel Tempering for Deep Posterior Approximation

The popular cosine learning rate is a special case of non-reversible parallel tempering.

 

ICML 2023

Modeling Temporal Data as Continuous Functions With Stochastic Process Diffusion

Here we augment generative diffusion with stochastic processes, which allows us to model irregular time sequences. 

 

ICML 2023

Provably Convergent Schrödinger Bridge With Applications to Probabilistic Time Series Imputation

In this paper, we propose a time series imputation and prediction model based on conditional generative AI and Schrodinger bridge SDE. 

 

AISTATS 2023

Risk Bounds for Aleatoric Uncertainty Recovery

We provide risk bounds for learning the data dependent variance function in a hetereoscedastic regression setting.

 

UAI 2023

Short-Term Temporal Dependency Detection Under Heterogeneous Event Dynamic With Hawkes Processes 

Here we propose a modified Hawkes process model that can better handle the unobserved background dynamics and artifacts.

 

UAI 2023

Inference and Sampling of Point Processes From Diffusion Excursions 

This paper proposes a new point process framework modeling arrival times through latent diffusion processes.

 

UAI 2023

In- or Out-of-Distribution Detection via Dual Divergence Estimation 

For Out-of-Distribution (OOD) detection, we propose a principled yet simple approach of (empirically) estimating KL-divergence, in its dual form, between the training and test sets. 

 

UAI 2023

Information Theoretic Clustering via Divergence Maximization Among Clusters 

For principled clustering with minimal apriori assumption, we propose to maximize the Kullback-Leibler divergence in its dual form between the underlying data distributions associated to clusters. 

 

NeurIPS 2023

Nearly Tight Bounds for Differentially Private Multiway Cut

We develop the first differentially private min s-t cut algorithm with tight approximation guarantees.

 

NeurIPS 2023

Topology-Aware Uncertainty for Image Segmentation

We propose a framework for modeling the uncertainty of the existence of topological structure.

 

ICLR 2023

Learning to Segment From Noisy Annotations: A Spatial Correction Approach

In this paper, we propose a novel Markov model for segmentation noisy annotations that encodes both spatial correlation and bias. 

 

International Journal of Forecasting

A Multi-Task Encoder-Dual-Decoder Framework for Mixed Frequency Data Prediction

This paper develops a neural network-based approach that handles forecasting and nowcasting in a unified fashion.

 

Short version is accepted at ICML 2023 Workshop

Reflected Schrödinger Bridge for Constrained Generative Modeling

We introduce the Reflected Schrodinger Bridge algorithm: an entropy-regularized optimal transport approach for generating data within diverse bounded domains.

2022

UAI 2022

Estimating Transfer Entropy Under Long-Ranged Dependencies

Here we estimate transfer entropy directly from conditional likelihoods computed in-sample using any timeseries forecaster trained per maximum likelihood principle. 

 

ICLR 2022

Interacting Contour Stochastic Gradient Langevin Dynamics

Interacting parallel stochastic gradient Langevin dynamics can be faster than the single long chain alternative with the same computational budget.

 

SIAM Journal on Optimization 2022

General Feasibility Bounds for Sample Average Approximation via Vapnik-Chervonenkis Dimension

This work improves the current understanding regarding the feasibility of sample average approximation solutions when solving stochastic programming problems without recourse.

 

UAI 2022

Stability of SGD: Tightness Analysis and Improved Bounds 

We construct a lower bound to examine the tightness of the existing theoretical results in the literature on generalization of Stochastic Gradient Descent.

 

IROS 2022

Scalable Safety-Critical Policy Evaluation With Accelerated Rare Event Sampling

We propose a policy evaluation based on adaptive sampling techniques for agents to evaluate and act on rare event, especially regarding safety constraints.

 

AI STATS 2022

A Manifold View of Adversarial Risk

We propose a new angle for decomposing adversarial risk into two orthogonal terms. Such a decomposition provides new directions for improving the robustness of a model.

2021

NeurIPS 2021

Topological Detection of Trojaned Neural Networks 

A trojaned neural network has a Trojan behavior related shortcut in its neuron graph. Such structural information is extracted with tools from algebraic topology.

2020

COLING 2020, Workshop on Financial Narrative Processing and MultiLing Financial Summarization

Information Extraction From Federal Open Market Committee Statements

We present a novel approach to unsupervised information extraction by identifying and extracting relevant concept-value pairs from textual data.