🏢 UC San Diego

VLG-CBM enhances concept bottleneck models with vision-language guidance for faithful interpretability and improved accuracy.

UniMTS, a novel unified pre-training model, achieves state-of-the-art performance in motion time series classification by generalizing across diverse device factors and activities.

TAKFL, a novel federated learning framework, tackles device heterogeneity by independently distilling knowledge from diverse devices and integrating it adaptively, achieving state-of-the-art performan…

Leveraging symmetry in automated equation discovery improves accuracy and simplicity of learned governing equations, enhancing robustness against noise and achieving higher success rates across divers…

SpatialRGPT enhances Vision-Language Models’ spatial reasoning by integrating 3D scene graphs and depth information, achieving significant performance gains on spatial reasoning tasks.

SharpBalance, a novel training approach, effectively improves deep ensemble performance by addressing the sharpness-diversity trade-off, leading to significant improvements in both in-distribution and…

This paper presents the first replicable uniformity tester with nearly linear dependence on the replicability parameter, enhancing the reliability of scientific studies using distribution testing algo…

RA-PbRL introduces a provably efficient algorithm for risk-aware preference-based reinforcement learning, addressing the limitations of existing risk-neutral methods in applications demanding heighten…

New APC metric precisely quantifies & optimizes global faithfulness in persona-driven role-playing, offering a fine-grained, explainable evaluation and improving AI character consistency.

One data point per class suffices for efficient and provable dataset distillation in kernel ridge regression, significantly reducing computational costs.

Online Label Shift adaptation with Online Feature Updates (OLS-OFU) significantly boosts online label shift adaptation by dynamically refining feature extractors using self-supervised learning, achiev…

The 1-nearest neighbor rule achieves online consistency under surprisingly broad conditions: measurable label functions and mild assumptions on instance generation in doubling metric spaces.

New algorithms achieve near-optimal differentially private U-statistic estimation, significantly improving accuracy over existing methods.

Transformers’ in-context learning mirrors human episodic memory, with specific attention heads acting like the brain’s contextual maintenance and retrieval system.

ZORRO: A new method for learning linear models from uncertain data, providing sound over-approximations of all possible models and prediction ranges.

IsoNet++ iteratively refines subgraph matching via early interaction GNNs and node-pair partner interactions, significantly boosting graph retrieval accuracy.

This paper proves that inexact ALMs applied to SDPs achieve linear convergence for both primal and dual iterates, contingent solely on strict complementarity and a bounded solution set, thus resol…

HYSYNTH: A hybrid approach uses LLMs to create context-free surrogate models that guide efficient program synthesis, outperforming LLMs alone and existing synthesizers across multiple domains.

Learning to rank request outputs improves LLM scheduling, resulting in 2.8x lower chatbot latency and 6.5x higher synthetic data generation throughput.

Generative models often produce invalid outputs; this work shows that ensuring validity is easier than expected when using log-loss and carefully selecting model classes and data distributions.