🏢 UC Los Angeles

SMALLTOLARGE (S2L) revolutionizes large language model (LLM) fine-tuning by using a small model to summarize training loss trajectories, enabling efficient data selection for larger models.

SimGen: Simulator-conditioned driving scene generation, uses a novel cascade diffusion pipeline to generate diverse driving scenes by mixing real-world and simulator data, addressing Sim2Real gaps.

IDA, a novel intervention assistance, dynamically shares control between human and AI copilots by intervening only when the AI’s action is superior across all goals, maximizing performance and preserv…

Self-Play Fine-Tuning (SPIN-Diffusion) revolutionizes diffusion model training, achieving superior text-to-image results with less data via iterative self-improvement, surpassing supervised and RLHF m…

Stormer, a simple transformer model, achieves state-of-the-art medium-range weather forecasting accuracy by using weather-specific embedding, randomized dynamics forecasting, and a pressure-weighted l…

SAFEWORLD: a new benchmark reveals and fixes LLMs’ struggle with diverse safety standards.

Reciprocators: AI agents that learn to cooperate by reciprocating influence, achieving prosocial outcomes in complex scenarios.

PUREGEN uses generative model dynamics to purify poisoned training data, providing a universal, effective, and efficient train-time defense against various data poisoning attacks.

LLMs’ in-context learning, though effective, exhibits surprisingly irregular decision boundaries, hindering generalization; this paper reveals this issue and proposes methods to improve smoothness via…

Pre-trained LLMs surprisingly use Fourier features to perform addition, with MLP layers approximating magnitude and attention layers handling modular arithmetic; this mechanism requires pre-training.

TREAT: a novel framework boosting dynamical systems modeling accuracy by enforcing Time-Reversal Symmetry (TRS) via a regularization term. High-precision modeling is achieved across diverse systems, …

Non-Euclidean Mixture Model (NMM-GNN) outperforms existing methods by using spherical and hyperbolic spaces to model homophily and social influence in social network embedding, improving link predicti…

Matryoshka Query Transformer (MQT) empowers large vision-language models with flexible visual token encoding, drastically reducing inference costs while maintaining high accuracy across multiple bench…

Sign Stochastic Gradient Descent (SGD) achieves optimal sample complexity for solving k-sparse parity problems, matching Statistical Query lower bounds.

Latent Plan Transformer (LPT) solves long-term planning challenges in reinforcement learning by using latent variables to connect trajectory generation with final returns, achieving competitive result…

Deep metric learning and Coreset integration enables efficient slice-based active learning for 3D medical segmentation, surpassing existing methods in performance with low annotation budgets.

GraphVis boosts LLMs by visualizing knowledge graphs, improving accuracy in textual and visual question answering.

ICTM efficiently solves linear inverse problems using flow priors by iteratively optimizing local MAP objectives, outperforming other flow-based methods.

Accelerated discrete diffusion model sampling is achieved via novel discrete non-Markov diffusion models (DNDM) with predetermined transition times, enabling a training-free algorithm that significant…

Self-Training on Image Comprehension (STIC) significantly boosts Large Vision Language Model (LVLM) performance using unlabeled image data. STIC generates a preference dataset for image descriptions …