Junhyeok Lee

Motif 3 achieves unprecedented efficiency and performance in language modeling by leveraging a novel Mixture-of-Experts architecture that activates only a fraction of its parameters per token.

ProPS can generate speaker embeddings that accurately reflect complex attributes from simple natural language prompts, revolutionizing how we synthesize speaker identities.

ST-MoME achieves unprecedented accuracy in synthesizing DCE-MRI maps from incomplete data, outperforming existing methods in clinically critical tumor regions.

TensorLDM achieves near-ground-truth diffusion tensor reconstruction with a remarkable SPD-violation rate of just 1.54%, setting a new standard for anatomical consistency in DTI.

Multilingual medical retrieval systems face a staggering performance drop, with scores plummeting from 0.818 in English to just 0.056 in Japanese, exposing a critical gap in current benchmarks.

SPARC achieves superior control performance in robotic applications by dynamically allocating bitrate based on the task relevance of visual information, outperforming traditional codecs.

Architectural specialization and an efficient training approach enable Motif-Video 2B to outperform larger models with a fraction of the resources.

Current medical vision-language models can't explain medical images to patients, but MedLayBench-V offers a way to fix that.

LLM-assisted scientific writing is producing more confident but homogenized prose, as evidenced by a 23% decline in hedging in the post-LLM era.

Quantifying uncertainty in physics-informed neural networks for medical imaging boosts accuracy and reliability, leading to better stroke assessment.