Jaegul Choo

Negative transfer in cross-domain recommendations can be effectively mitigated with a dual-expert LLM approach that enhances item-level understanding.

Achieving over 90% accuracy retention with 50% pruning, DVBP + OB²C outperforms traditional methods by leveraging noise-filtered neuron selection and optimal weight updates.

ReDesign not only recovers editable design files from images but does so with unprecedented accuracy and flexibility, setting a new standard for design workflows.

Disentangling the intricate factors affecting inference speed reveals that effective acceleration of dLLMs hinges on specialized techniques rather than mere parallel generation.

Robot-centric pointmaps can drastically improve action prediction in VLA models, especially when faced with diverse camera viewpoints.

3D HAMSTER achieves superior robotic manipulation performance by directly predicting 3D trajectories, eliminating the distortions caused by 2D guidance.

Zero-shot transfer of a refined RL policy boosts manipulation success rates from 42% to 76% on real robots, showcasing a breakthrough in sim-to-real applications.

Membrane achieves the highest safety performance against evolving jailbreaks while keeping benign refusals remarkably low, revolutionizing LLM defense strategies.

Off-policy RL can now train in minutes instead of hours for sim-to-real robot control, thanks to a new algorithm that borrows scaling laws from supervised learning.

Unlock the potential of full-duplex speech language models with Sommelier, a new open-source pipeline that tackles the messy reality of multi-speaker conversations.

Injecting noise estimates layer-by-layer into diffusion models dramatically improves speech enhancement in complex, real-world conditions where single-point injection fails.