Jongwon Park

Heehoon Kim, Jaehwan Lee, Taejeoung Kim et al.

The rapid growth of large language models is driving organizations to expand their GPU clusters, often with GPUs from multiple vendors. However, current deep learning frameworks lack support for collective communication across heterogeneous GPUs, leading to inefficiency and higher costs. We present HetCCL, a collective communication library that unifies vendor-specific backends and enables RDMA-based communication across GPUs without requiring driver modifications. HetCCL introduces two novel mechanisms that enable cross-vendor communication while leveraging optimized vendor libraries, NVIDIA NCCL and AMD RCCL. Evaluations on a multi-vendor GPU cluster show that HetCCL matches NCCL and RCCL performance in homogeneous setups while uniquely scaling in heterogeneous environments, enabling practical, high-performance training with both NVIDIA and AMD GPUs without changes to existing deep learning applications.

Daeun Lee, Jongwon Park, Jinkyu Kim

An autonomous driving system requires a 3D object detector, which must perceive all present road agents reliably to navigate an environment safely. However, real-world driving datasets often suffer from the problem of data imbalance, which causes difficulties in training a model that works well across all classes, resulting in an undesired imbalanced sub-optimal performance. In this work, we propose a method to address this data imbalance problem. Our method consists of two main components: (i) a LiDAR-based 3D object detector with per-class multiple detection heads where losses from each head are modified by dynamic weight average to be balanced. (ii) Contextual ground truth (GT) sampling, where we improve conventional GT sampling techniques by leveraging semantic information to augment point cloud with sampled ground truth GT objects. Our experiment with KITTI and nuScenes datasets confirms our proposed method's effectiveness in dealing with the data imbalance problem, producing better detection accuracy compared to existing approaches.

Jongwon Park, Heesoo Jung, Hogun Park

Recent Self-Supervised Learning (SSL) methods encapsulating relational information via masking in Graph Neural Networks (GNNs) have shown promising performance. However, most existing approaches rely on random masking strategies in either feature or graph space, which may fail to capture task-relevant information fully. We posit that this limitation stems from an inability to achieve minimum redundancy between masked and unmasked components while ensuring maximum relevance of both to potential downstream tasks. Conditional Independence (CI) inherently satisfies the minimum redundancy and maximum relevance criteria, but its application typically requires access to downstream labels. To address this challenge, we introduce CIMAGE, a novel approach that leverages Conditional Independence to guide an effective masking strategy within the latent space. CIMAGE utilizes CI-aware latent factor decomposition to generate two distinct contexts, leveraging high-confidence pseudo-labels derived from unsupervised graph clustering. In this framework, the pretext task involves reconstructing the masked second context solely from the information provided by the first context. Our theoretical analysis further supports the superiority of CIMAGE's novel CI-aware masking method by demonstrating that the learned embedding exhibits approximate linear separability, which enables accurate predictions for the downstream task. Comprehensive evaluations across diverse graph benchmarks illustrate the advantage of CIMAGE, with notably higher average rankings on node classification and link prediction tasks. Notably, our proposed model highlights the under-explored potential of CI in enhancing graph SSL methodologies and offers enriched insights for effective graph representation learning.

Jongwon Park, Minku Kang, Wooseok Sim et al.

Federated Learning (FL) is gaining prominence in machine learning as privacy concerns grow. This paradigm allows each client (e.g., an individual online store) to train a recommendation model locally while sharing only model updates, without exposing the raw interaction logs to a central server, thereby preserving privacy in a decentralized environment. Nonetheless, most existing FL-based recommender systems still rely solely on each client's private data, despite the abundance of publicly available datasets that could be leveraged to enrich local training; this potential remains largely underexplored. To this end, we consider a realistic scenario wherein a large shopping platform collaborates with multiple small online stores to build a global recommender system. The platform possesses global data, such as shareable user and item lists, while each store holds a portion of interaction data privately (or locally). Although integrating global data can help mitigate the limitations of sparse and biased clients' local data, it also introduces additional challenges: simply combining all global interactions can amplify noise and irrelevant patterns, worsening personalization and increasing computational costs. To address these challenges, we propose FedGDVE, which selectively augments each client's local graph with semantically aligned samples from the global dataset. FedGDVE employs: (i) a pre-trained graph encoder to extract global structural features, (ii) a local valid predictor to assess client-specific relevance, (iii) a reinforcement-learning-based probability estimator to filter and sample only the most pertinent global interactions. FedGDVE improves performance by up to 34.86% on recognized benchmarks in FL environments.

Inhyo Lee, Junhyeong Lee, Jongwon Park et al.

Double perovskites (DPs) are promising candidates for sustainable energy technologies due to their compositional tunability and compatibility with low-energy fabrication, yet their vast design space poses a major challenge for conditional materials discovery. This work introduces a multi-agent, text gradient-driven framework that performs DP composition generation under natural-language conditions by integrating three complementary feedback sources: LLM-based self-evaluation, DP-specific domain knowledge-informed feedback, and ML surrogate-based feedback. Analogous to how knowledge-informed machine learning improves the reliability of conventional data-driven models, our framework incorporates domain-informed text gradients to guide the generative process toward physically meaningful regions of the DP composition space. Systematic comparison of three incremental configurations, (i) pure LLM generation, (ii) LLM generation with LLM reasoning-based feedback, and (iii) LLM generation with domain knowledge-guided feedback, shows that iterative guidance from knowledge-informed gradients improves stability-condition satisfaction without additional training data, achieving over 98% compositional validity and up to 54% stable or metastable candidates, surpassing both the LLM-only baseline (43%) and prior GAN-based results (27%). Analyses of ML-based gradients further reveal that they enhance performance in in-distribution (ID) regions but become unreliable in out-of-distribution (OOD) regimes. Overall, this work provides the first systematic analysis of multi-agent, knowledge-guided text gradients for DP discovery and establishes a generalizable blueprint for MAS-driven generative materials design aimed at advancing sustainable technologies.

Jinpyo Kim, Gyeongje Cho, Chanwoo Park et al.

Since state-of-the-art LLMs often underperform in languages other than English or Chinese, improving the capability of LLMs in new languages has become an essential task. Moreover, LLMs' entire end-to-end training process remains largely unknown to the public due to proprietary reasons, technical complexity, inconsistent documentation, and ethical considerations. The complete picture remains a closely guarded secret within the industry. This paper presents methods to adapt an existing English-based LLM to Korean in a low-budget scenario. We describe the entire end-to-end process: collecting Korean datasets, preprocessing the data, training the model, creating downstream benchmarks, and conducting evaluations. The evaluation results indicate that our method can effectively and cost-efficiently add new language capabilities to existing LLMs. Our new bilingual models, Thunder-LLM and Thunder-LLM-Ins, achieve superior Korean performance compared to state-of-the-art models while utilizing minimal data and computational resources. We share our comprehensive experience and make the code publicly available.