Jiwon Kim

A single benchmark now enables city-scale, multimodal OD flow forecasting across six modes and sub-district granularity, revealing that scaling spatio-temporal models and naive joint training often fail.

EXAONE Tabular outperforms tuned ensembles and larger models, achieving top rankings in multiple benchmarks while drastically reducing inference costs.

The 1-block STGCN variant achieves superior performance with significantly lower computational costs, challenging the assumption that deeper architectures are always better for traffic prediction.

LLMs may mimic human responses at the surface level but often miss the deeper distributional nuances, revealing a critical flaw in conventional evaluation methods.

Open-source LLMs, trained with Reddit community feedback, can match the performance of proprietary models in generating supportive mental health responses, offering a privacy-preserving alternative.

Content moderation systems, designed to protect users, may inadvertently cripple LLMs' ability to function as effective therapists by censoring crucial topics in real therapy sessions.

Achieve faster, smaller medical image compression without sacrificing diagnostic quality by simply throwing away the background *before* using your favorite codec.