🏢 Purdue University

RLbreaker uses deep reinforcement learning to efficiently create highly effective jailbreaking prompts, outperforming existing methods against multiple state-of-the-art LLMs and defenses.

AI explanations can be subtly manipulated to influence human decisions, highlighting the urgent need for more robust and ethical AI explanation design.

Active learning’s optimal rates are completely characterized, resolving an open problem and providing new algorithms achieving exponential and sublinear rates depending on combinatorial complexity mea…

Unified Covariate Adjustment (UCA) offers a scalable, doubly robust estimator for a wide array of causal estimands beyond standard methods.

ProRec, a novel framework, bridges the binary-source semantic gap by using a binary-source encoder-decoder model and LLMs, achieving significant improvements in zero-shot binary summarization and func…

Soft Superpixel Neighborhood Attention (SNA) optimally denoises images by incorporating superpixel probabilities into an attention module, outperforming traditional methods.

Efficiently learn causal graphs from limited interventions using a novel Bayesian algorithm that outperforms existing methods and requires fewer experiments.

Learning optimal interventions in causal bandits with unknown causal graphs is now efficient; this paper identifies the minimal causal knowledge needed and offers a two-stage algorithm with sublinear …

OPEL: a novel optimal transport framework for procedure learning, significantly outperforms SOTA methods by aligning similar video frames and relaxing strict temporal assumptions.

Unbounded label spaces conquered! New algorithm achieves optimal mistake bounds in multiclass transductive online learning.

LLMDFA: A novel LLM-powered framework performs compilation-free and customizable dataflow analysis, achieving high accuracy in bug detection by decomposing the task into sub-problems and mitigating L…

LEDEX: A novel training framework significantly boosts LLMs’ code self-debugging by using automated data collection, supervised fine-tuning, and reinforcement learning, leading to more accurate code a…

First-ever sublinear regret & constraint violation bounds achieved for infinite horizon average reward CMDPs with general policy parametrization using a novel primal-dual policy gradient algorithm.

This paper presents novel theoretical frameworks and algorithms for learning from snapshots of discrete and continuous data streams, resolving key learnability challenges in online learning under cont…

This paper significantly improves our understanding of multiclass PAC learning by reducing the sample complexity gap and proposing two novel approaches to fully resolve the optimal sample complexity.

MTGC tackles multi-timescale model drift in hierarchical federated learning.

Deep neural networks surprisingly exhibit universal convergence in input salience, aligning more closely as model capacity increases, revealing valuable insights into model behavior and improving deep…

ACS: Automatic Cyclical Scheduling revolutionizes gradient-based discrete sampling by intelligently switching between exploration and exploitation phases to efficiently navigate complex multimodal dis…

FIARSE dynamically optimizes submodels in federated learning based on parameter importance, improving efficiency and global model accuracy.

DSPoser: A novel two-stage approach accurately estimates full-body pose from doubly sparse egocentric video data using masked autoencoders for temporal completion and conditional diffusion models for …