🏢 KAIST

Self-guidance boosts masked generative models’ image synthesis, achieving superior quality and diversity with fewer steps!

TrackIME enhances video point tracking by cleverly pruning the search space, resulting in improved accuracy and efficiency.

PixelCLIP: Open-vocabulary semantic segmentation without pixel-level labels! Leveraging unlabeled image masks from Vision Foundation Models and an online clustering algorithm, PixelCLIP achieves imp…

SyncTweedies: a zero-shot diffusion synchronization framework generates diverse visual content (images, panoramas, 3D textures) by synchronizing multiple diffusion processes without fine-tuning, demon…

Researchers extended stochastic optimal control theory to infinite-dimensional spaces, enabling the creation of diffusion bridges for generative modeling in function spaces, demonstrating applications…

Stochastic extragradient with flip-flop shuffling & anchoring achieves provably faster convergence in minimax optimization.

Train Neural ODEs without simulations, achieving high performance on regression and classification by using flow matching in the embedding space of data pairs.

RegQ: A novel regularized Q-learning algorithm ensures convergence with linear function approximation, solving a long-standing instability problem in reinforcement learning.

CutMix and Cutout data augmentation methods provably improve feature learning by enabling the network to learn rarer features and noise vectors more effectively.

LLMs boost tabular data prediction by generating optimized features via decision tree reasoning, outperforming existing methods.

MAC: Efficiently updates large language models (LLMs) using a memory of compressed contexts for improved real-time knowledge retention and adaptation.

Learn disentangled 3D object poses and transfer them between different object identities using a novel neural pose representation, boosting 3D shape generation!

MHCDIFF: a novel pipeline using multi-hypotheses conditioned point cloud diffusion for accurate 3D human reconstruction from occluded images, outperforming state-of-the-art methods.

f-RAG: A novel fragment-based molecular generation framework boosts drug discovery by combining retrieval augmentation with a generative model, enabling exploration beyond existing fragments and signi…

Bayesian Optimization Model Fusion (BOMF) significantly boosts language model fine-tuning by optimizing both loss and metrics through multi-objective Bayesian optimization, yielding considerable perfo…

DrilDICE robustly tackles covariate shift in offline imitation learning by using a stationary distribution correction and a distributionally robust objective, significantly improving performance.

Meteor: Mamba-based Traversal of Rationale achieves significant vision-language improvements by efficiently embedding multifaceted rationales in a large language model, without scaling the model or us…

Decoupled Token Embedding for Merging (DTEM) significantly improves Vision Transformer efficiency by using a decoupled embedding module for relaxed token merging, achieving consistent performance gain…

MuDI: a novel framework for multi-subject image personalization, effectively decoupling identities to prevent mixing using segmented subjects and a new evaluation metric.

LLMs’ factual knowledge acquisition during pretraining is surprisingly non-linear: more data doesn’t guarantee better knowledge retention, and forgetting follows a power law.