Soo Kyung Kim

Soo Kyung Kim, Kalai Ramea, Salva Rühling Cachay et al.

The availability of training data remains a significant obstacle for the implementation of machine learning in scientific applications. In particular, estimating how a system might respond to external forcings or perturbations requires specialized labeled data or targeted simulations, which may be computationally intensive to generate at scale. In this study, we propose a novel solution to this challenge by utilizing a principle from statistical physics known as the Fluctuation-Dissipation Theorem (FDT) to discover knowledge using an AI model that can rapidly produce scenarios for different external forcings. By leveraging FDT, we are able to extract information encoded in a large dataset produced by Earth System Models, which includes 8250 years of internal climate fluctuations, to estimate the climate system's response to forcings. Our model, AiBEDO, is capable of capturing the complex, multi-timescale effects of radiation perturbations on global and regional surface climate, allowing for a substantial acceleration of the exploration of the impacts of spatially-heterogenous climate forcers. To demonstrate the utility of AiBEDO, we use the example of a climate intervention technique called Marine Cloud Brightening, with the ultimate goal of optimizing the spatial pattern of cloud brightening to achieve regional climate targets and prevent known climate tipping points. While we showcase the effectiveness of our approach in the context of climate science, it is generally applicable to other scientific disciplines that are limited by the extensive computational demands of domain simulation models. Source code of AiBEDO framework is made available at https://github.com/kramea/kdd_aibedo. A sample dataset is made available at https://doi.org/10.5281/zenodo.7597027. Additional data available upon request.

Hyun Lee, Hyemin Jeong, Yejin Kim et al.

Modern multimodal large language models (MLLMs) typically keep the language model fixed and train a visual projector that maps the pixels into a sequence of tokens in its embedding space, so that images can be presented in essentially the same form as text. However, the language model has been optimized to operate on discrete, semantically meaningful tokens, while prevailing visual projectors transform an image into a long stream of continuous and highly correlated embeddings. This causes the visual tokens to behave differently from the word-like units that LLMs are originally trained to understand. We propose a novel Disentangled Visual Tokenization (DiVT) that clusters patch embeddings into coherent semantic units, so each token corresponds to a distinct visual concept instead of a rigid grid cell. DiVT further adapts its token budget to image complexity, providing an explicit accuracy-compute trade-off modifying neither the vision encoder nor the language model. Across diverse multimodal benchmarks, DiVT matches or surpasses baselines with significantly fewer visual tokens, demonstrating robustness under limited token budgets, significantly reducing memory cost and latency while making visual inputs more compatible with LLMs. Our code is available at https://github.com/snuviplab/DiVT.

Daeun Kyung, Hyunseung Chung, Seongsu Bae et al.

Doctor-patient consultations require multi-turn, context-aware communication tailored to diverse patient personas. Training or evaluating doctor LLMs in such settings requires realistic patient interaction systems. However, existing simulators often fail to reflect the full range of personas seen in clinical practice. To address this, we introduce PatientSim, a patient simulator that generates realistic and diverse patient personas for clinical scenarios, grounded in medical expertise. PatientSim operates using: 1) clinical profiles, including symptoms and medical history, derived from real-world data in the MIMIC-ED and MIMIC-IV datasets, and 2) personas defined by four axes: personality, language proficiency, medical history recall level, and cognitive confusion level, resulting in 37 unique combinations. We evaluate eight LLMs for factual accuracy and persona consistency. The top-performing open-source model, Llama 3.3 70B, is validated by four clinicians to confirm the robustness of our framework. As an open-source, customizable platform, PatientSim provides a reproducible and scalable solution that can be customized for specific training needs. Offering a privacy-compliant environment, it serves as a robust testbed for evaluating medical dialogue systems across diverse patient presentations and shows promise as an educational tool for healthcare. The code is available at https://github.com/dek924/PatientSim.