bio photo

Bosung Kim

she/her

View My GitHub Profile

Hi!👋

I’m a PhD candidate in Computer Science and Engineering at UC San Diego🌴☀️🏄🏻‍♀️🌊, advised by Prithviraj in the PEARLS Lab🧋.

My research focuses on AI agents🤖 that use long-context, multimodal information to reason and act in embodied environments. I’m particularly interested in enabling agents to complete long-horizon tasks and translating the reasoning capabilities of frontier models into effective robot actions.

I began my research journey in natural language processing and generative AI, especially in leveraging LLMs for knowledge extraction. Previously, I’ve interned at NVIDIA Research , Microsoft Research , AWS AI Lab , and Megagon Labs.

[CV] [Google scholar] [X]

News

Publications

GPT-6 Astra connected to four robot tasks: claw machine, stacking cups, folding cloth, and closing a lid.

Bridging Frontier Reasoning and Robot Execution: From Autonomous Demonstration Generation to Dense Language Supervision
[paper] [project page]
Bosung Kim, Alexander Trevithick, Ruiyi Wang, Prithviraj Ammanabrolu
arXiv preprint, 2026

DeMiAn highlights: dense language annotations for robot learning, with a montage of human demonstrations.

How to Instruct Your Robot: Dense Language Annotations Power Robot Policy Learning
[paper] [project page]
Bosung Kim*, Ruiyi Wang*, David Acuna, Jaehun Jung, Alexander Trevithick, Brandon Cui, Yejin Choi, Prithviraj Ammanabrolu
NeurIPS, 2026 · * equal contribution

DECADE: Evaluating Evolving Long-Term Memory in Personalized Conversational AI Assistants
Bosung Kim, Hao Cheng, Yuan Gao, Prithviraj Ammanabrolu, Michel Galley, Rachel Rong, Jianfeng Gao
Under review, 2026

First-person exploration of a household environment in the long-context ∞-THOR benchmark.

Beyond Needle(s) in the Embodied Haystack: Environment, Architecture, and Training Considerations for Long Context Reasoning
[paper] [project page]
Bosung Kim, Prithviraj Ammanabrolu
CVPR Workshop FMEA, 2025

SYMPTOMIFY: Transforming Symptom Annotations with Language Model Knowledge Harvesting
[paper]
Bosung Kim, Ndapa Nakashole
EMNLP Findings, 2023

ZETT pipeline: fine-tuning and zero-shot relation extraction by template infilling, scoring, and ranking.

Zero-shot Triplet Extraction by Template Infilling
[paper] [project page]
Bosung Kim, Hayate Iso, Nikita Bhutani, Estevam Hruschka, Ndapa Nakashole, Tom Mitchell
IJCNLP-AACL, 2023

Chride at SemEval-2023 Task 10: Fine-tuned DebertaV3 on Detection of Online Sexism with Hierarchical Loss
[paper]
Letian Peng, Bosung Kim
ACL Workshop SemEval, 2023

Data Augmentation for Rare Symptoms in Vaccine Side-Effect Detection
[paper]
Bosung Kim, Ndapa Nakashole
ACL Workshop BioNLP, 2022

Dictionary graph construction, GNN pre-training, query encoder training, and jargon-word inference.

Query Reformulation for Descriptive Queries of Jargon Words Using a Knowledge Graph based on a Dictionary
[paper]
Bosung Kim, Hyewon Choi, Haeun Yu, Youngjoong Ko
CIKM, 2021

Commonsense knowledge augmentation through candidate generation, adversarial validity scoring, and augmentation.

Commonsense Knowledge Augmentation for Low-Resource Languages via Adversarial Learning
[paper]
Bosung Kim*, Juae Kim*, Youngjoong Ko, Jungyun Seo
AAAI, 2021 · * equal contribution

Shared BERT encoder with link prediction, relation prediction, and relevance ranking tasks.

Multi-task learning for Knowledge Graph Completion using Pre-trained Language Models
[paper]
Bosung Kim, Taesuk Hong, Youngjoong Ko, Jungyun Seo
COLING, 2020

Journal Articles

Figure 4: class-conditional output distributions for positive and negative examples.

Novel Regularization Method for the Class Imbalance Problem
[paper]
Bosung Kim, Youngjoong Ko, Jungyun Seo
Expert Systems with Applications, 2022