🏢 Seoul National University

Hassle-Free Textual Training (HFTT) uses only textual data to effectively remove unwanted visual data from AI training datasets, significantly reducing human annotation needs.

SRCPO: a novel spectral risk measure-constrained RL algorithm guaranteeing convergence to a global optimum, outperforming existing methods in continuous control tasks.

Self-guided MAE boosts self-supervised learning by intelligently masking image patches based on internal clustering patterns, dramatically accelerating training without external data.

ConFrag, a novel approach to noisy label regression, leverages contrastive fragmentation and neighborhood agreement to select clean samples, significantly outperforming state-of-the-art baselines on s…

First provably efficient randomized RL algorithms using multinomial logistic function approximation are introduced, achieving superior performance and constant-time computational cost.

Novel algorithms stabilize multi-server queueing systems with unknown service rates, achieving sublinear regret by learning server preferences via preference feedback.

Unified Spoken Dialog Model (USDM) directly generates coherent spoken responses with natural prosody, surpassing cascaded baselines and enhancing natural conversation in speech-enabled LLMs.

This paper presents OFU-MNL+, a constant-time algorithm achieving nearly minimax optimal regret for contextual multinomial logistic bandits, closing the gap between existing upper and lower bounds.

BinaryMoS: a novel token-adaptive binarization method that boosts LLM accuracy and efficiency by dynamically merging multiple scaling experts for each token.

DPR, a novel bias mitigation method, robustly improves model performance by leveraging disagreement probability without needing bias labels, achieving state-of-the-art results.

Spectral Attention boosts long-range dependency capture in time series forecasting, achieving state-of-the-art results across various models and datasets.

Linear ensemble sampling achieves a state-of-the-art regret bound of Õ(d³/²√T) with a logarithmic ensemble size, closing the theory-practice gap in linear bandit algorithms.

This paper introduces a novel gradient-free decoder inversion method for latent diffusion models, improving efficiency and memory usage compared to existing gradient-based methods. The method is theo…

GIN, a novel approximate Bayesian inference algorithm, efficiently handles nonlinear state-space models with high-dimensional, noisy observations by disentangling observation and dynamics. Achieving l…

FIFO-Diffusion generates infinitely long, high-quality videos from text prompts using a pretrained model, solving the challenge of long video generation without retraining.

FedAvP enhances federated learning’s privacy by sharing only augmentation policies, improving performance in diverse settings.

DropBP: Accelerate LLM fine-tuning by 44% while preserving accuracy!

Cross-Attention-only Time Series Transformer (CATS) outperforms existing models by removing self-attention, improving long-term forecasting accuracy, and reducing computational cost.

Adaptive algorithm achieves tight regret bounds for infinitely many-armed bandits under generalized rotting constraints, addressing the challenge of decreasing rewards over time.

Regret-Guided Diffusion Models enhance unsupervised environment design by generating challenging, diverse training environments that improve agent robustness and zero-shot generalization.