🏢 University of Texas at Austin

MatFormer: Train one universal model, extract hundreds of accurate submodels for elastic inference!

LOFIT: Localized fine-tuning boosts LLMs’ performance by selectively training only a small subset of attention heads, achieving comparable accuracy to other methods while using significantly fewer par…

LightGaussian achieves 15x compression of 3D Gaussian scene representations, boosting rendering speed to 200+ FPS while maintaining visual quality, solving storage and efficiency issues in real-time n…

Proper learning of noisy halfspaces with margins is achievable with sample complexity matching random classification noise, defying prior expectations.

Softmax attention in transformers adapts its attention window to function Lipschitzness and noise, enabling efficient in-context learning.

Training high-quality diffusion models efficiently is now possible, thanks to novel sample complexity bounds improving exponentially on previous work.

Researchers can now pinpoint the sources of data shifts in complex linear causal systems using a new algorithm, even with limited perfect interventions, opening exciting possibilities for causal disco…

HydraLoRA: Asymmetric LoRA boosts LLM fine-tuning efficiency by sharing parameters across tasks while specializing others, outperforming existing methods.

HOI-Swap: a novel diffusion model flawlessly swaps objects in videos while intelligently preserving natural hand interactions, producing high-quality edits.

Hierarchical Hybrid Sliced Wasserstein (H2SW) solves the challenge of comparing complex, heterogeneous joint distributions by introducing novel slicing operators, leading to a scalable and statistical…

HiCS-FL: A novel federated learning client sampling method that leverages data heterogeneity for faster, more efficient global model training in non-IID settings.

This paper introduces a rate-distortion framework for prompt compression in LLMs, bridging the gap between existing methods and optimal performance. By formulating prompt compression as a linear progr…

Ms-PoE, a simple plug-and-play positional encoding, significantly improves LLMs’ ability to utilize long contexts by mitigating the ’lost-in-the-middle’ problem and enhancing the capacity to capture i…

EVA: a novel method generates expressive 3D Gaussian human avatars from monocular RGB videos, excelling in detailed hand and facial expressions via context-aware density control and improved SMPL-X al…

Provably efficient algorithms for learning with distribution shift are introduced, generalizing and improving prior work by achieving near-optimal error rates and offering universal learners for large…

Dynamic Model Predictive Shielding (DMPS) ensures provably safe reinforcement learning by dynamically optimizing reinforcement learning objectives while maintaining provable safety, achieving higher r…

DUSDi: A novel method for learning disentangled skills in unsupervised reinforcement learning, enabling efficient reuse for diverse downstream tasks.

DUPLEX: a novel RL method trains diverse, near-optimal policies in complex, dynamic environments by explicitly maximizing policy diversity using successor features. It outperforms existing methods in…

Hybrid LLM & rule-based system accurately detects costly smart contract bugs!

Distributed Lion: Training large AI models efficiently by communicating only binary or low-precision vectors between workers and a server, significantly reducing communication costs and maintaining co…