🏢 University of Illinois Urbana-Champaign

Researchers developed Spherical Convolutional Wasserstein Distance (SCWD) to more accurately validate climate models by considering spatial variability and local distributional differences.

SnapKV: Slashing LLM memory usage & boosting speed via smart KV cache compression!

This work presents a sharper analysis of sketching for distributed deep learning, eliminating the problematic dependence on ambient dimension in convergence analysis and proving ambient dimension-inde…

SelfCodeAlign is a novel self-alignment method for code generation LLMs that surpasses existing methods by avoiding reliance on expensive human annotation or proprietary LLMs. The method achieves thi…

BICCOS: Scalable neural network verification via branch-and-bound inferred cutting planes.

Robust Prompt Optimization (RPO) creates robust LLM defenses against jailbreaking attacks by optimizing a transferable suffix, achieving state-of-the-art robustness.

RABBit: A novel Branch-and-Bound verifier for precise relational verification of Deep Neural Networks, achieving substantial precision gains over current state-of-the-art baselines.

Boost online finetuning of Decision Transformers by adding TD3 gradients, especially when pretrained with low-reward data.

Regularizing hidden states improves reward model generalization in RLHF for LLMs, boosting accuracy and mitigating over-optimization.

RAMP: A novel training framework significantly boosts DNN robustness against diverse adversarial attacks by mitigating accuracy-robustness tradeoffs and improving generalization.

ProEdit: High-quality 3D scene editing via progressive subtask decomposition.

PageRank Bandits (PRB) revolutionizes link prediction by framing it as a sequential decision-making problem, thus enabling the system to adapt to evolving data. Combining contextual bandits with PageR…

This paper proposes a novel, reward-free RLHF framework using a general preference oracle, surpassing existing reward-based approaches in efficiency and generalizability.

Tree-structured probabilistic circuits are surprisingly efficient: this paper proves a quasi-polynomial upper bound on their size, showing they’re almost as expressive as more complex DAG structures.

This paper tackles the ‘curse of horizon’ in off-policy evaluation for partially observable Markov decision processes (POMDPs) by proposing novel coverage assumptions, enabling polynomial estimation e…

Model-free policy gradient methods using occupancy functions are developed for online and offline RL, achieving computational efficiency and handling arbitrary data distributions.

Adaptive greedy algorithms significantly improve the accuracy of identifying the most influential subset of training data, overcoming limitations of existing methods that fail to capture complex inter…

Lexicon3D: a first comprehensive study probing diverse visual foundation models for superior 3D scene understanding, revealing that unsupervised image models outperform others across various tasks.

InterDreamer: Zero-shot text-guided 3D human-object interaction generation without paired data, achieved via decoupled semantic and dynamic modeling, using LLMs and a physics-based world model.

Fine-grained control is added to generative models for IoT sensing data augmentation, tailoring synthetic data to specific application needs by leveraging domain expertise and statistical metrics of s…