Papers by Soeun Kim
IFCap: Image-like Retrieval and Frequency-based Entity Filtering for Zero-shot Captioning (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing text-only training methods overlook the modality gap between using text data during training and employing images during inference. |
| Approach: | They propose a novel approach that aligns text features with visually relevant features to mitigate the modality gap between using text data during training and employing images during inference. |
| Outcome: | The proposed method outperforms the state-of-the-art methods in image captioning and video captioning by a significant margin compared to training with text data. |
Assigning Distinct Roles to Quantized and Low-Rank Matrices Toward Optimal Weight Decomposition (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing joint optimization methods prioritize one component at the expense of the other, resulting in suboptimal decompositions that fail to leverage each component’s unique strengths. |
| Approach: | They introduce Outlier-Driven Low-Rank Initialization (ODLRI) which assigns low-rank components the specific role of capturing activation-sensitive weights. |
| Outcome: | Experiments on Llama2 (7B, 13B, 70B, and Mistral-7B) and Llma3-8B show that incorporating ODLRI into the joint optimization framework reduces activation-aware error, minimizes quantization scale, and improves perplexity and zero-shot accuracy in low-bit settings. |
SPeCtrum: A Grounded Framework for Multidimensional Identity Representation in LLM-Based Agent (2025.naacl-long)
Copied to clipboard
Keyeun Lee, Seo Hyeong Kim, Seolhee Lee, Jinsu Eun, Yena Ko, Hayeon Jeon, Esther Hehsun Kim, Seonghye Cho, Soeun Yang, Eun-mee Kim, Hajin Lim
| Challenge: | Existing methods for simulating individual identities oversimplify human complexity, leading to incomplete or flattened representations. |
| Approach: | They propose a framework for constructing authentic LLM agent personas by incorporating an individual’s multidimensional self-concept. |
| Outcome: | The framework integrates Social Identity (S), Personal Identity (P), and Personal Life Context (C) components, each contributing distinct yet interconnected aspects of identity. |
DUSK: Do Not Unlearn Shared Knowledge (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work suggests that machine learning models are indistinguishable from models trained on retain sets. |
| Approach: | They propose a benchmark to evaluate machine unlearning under realistic knowledge overlap . they construct documents containing both shared and unique knowledge . |
| Outcome: | The proposed model is indistinguishable from a model retrained on the retain set while only forget-specific content is removed. |