Papers by Yongil Kim
SWITCH: Studying with Teacher for Knowledge Distillation of Large Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Knowledge Distillation (KD) has emerged as a popular method for compressing large language models due to high inference costs and memory requirements. |
| Approach: | They propose a method that integrates the teacher model during the student's sequence generation to reduce misguidance from the teacher. |
| Outcome: | Experiments on three model families and five instruction-following datasets show that SWITCH surpasses traditional methods, especially in the generation of long sequential data. |
Modality Alignment between Deep Representations for Effective Video-and-Language Learning (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing Video-and-Language models do not take into account the different characteristics of video and text representations. |
| Approach: | They propose a method that exploits Centered Kernel Alignment (CKA) to enhance cross-modality attention by combining multiple modalities. |
| Outcome: | The proposed method outperforms conventional multi-modal methods significantly on video QA tasks with +3.57% accuracy increment compared to the baseline in a popular benchmark dataset. |
PR-MCS: Perturbation Robust Metric for MultiLingual Image Captioning (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing image captioning metrics are vulnerable to lexical perturbations, but they are not robust to such perturbations. |
| Approach: | They propose a perturbation-robust multilingual CLIPScore which is a reference-free image captioning metric for multiple languages. |
| Outcome: | The proposed metric outperforms baseline metrics in capturing lexical noise of all various perturbation types in all five languages while maintaining a strong correlation with human judgments. |
Don’t Judge Code by Its Cover: Exploring Biases in LLM Judges for Code Evaluation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly used as evaluators for code evaluation tasks . however, whether they can handle superficial variations remains unclear . |
| Approach: | They define six types of potential biases in code evaluation and reveal their impact on LLM judges. |
| Outcome: | The proposed method can be used to evaluate semantically equivalent code with superficial variations without reference implementations. |
Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation (2025.emnlp-main)
Copied to clipboard
| Challenge: | a recent study has found that large vision–language models are vulnerable to visual biases that inflate scores without altering semantic content. |
| Approach: | They propose a novel meta-evaluation benchmark that exhibits diverse score distributions. |
| Outcome: | The proposed model exhibits vulnerability across all domains, and combines multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. |
MP2D: An Automated Topic Shift Dialogue Generation Framework Leveraging Knowledge Graphs (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to manage topic shifts within on-topic dialogues are limited in their ability to generate training datasets. |
| Approach: | They propose a data generation framework that automatically generates conversational question-answering datasets with natural topic transitions by leveraging relationships between entities in a knowledge graph. |
| Outcome: | The proposed framework generates conversational question-answering datasets with natural topic transitions and proves its effectiveness in generating dialogues with topic shifts. |
Dialogizer: Context-aware Conversational-QA Dataset Generation from Textual Sources (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing dialog inpainting methods generate ConvQA datasets with low contextual relevance due to insufficient learning of question-answer alignment. |
| Approach: | They propose a dialog inpainting method that generates ConvQA datasets from documents . they propose re-ranking tasks and a framework that generate contextually relevant questions . |
| Outcome: | The proposed framework generates ConvQA datasets with high contextual relevance from textual sources. |
Ko-LongRAG: A Korean Long-Context RAG Benchmark Built with a Retrieval-Free Approach (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Existing benchmarks for long-context RAG focus primarily on English . low-resource languages lack comprehensive evaluation frameworks limiting their progress in retrieval-based tasks. |
| Approach: | Ko-LongRAG is the first Korean long-context RAG benchmark . it adopts a retrieval-free approach designed around Specialized Content Knowledge (SCK) o1 model achieves the highest performance among proprietary models, while EXAONE 3.5 leads among open-sourced models . |
| Outcome: | the benchmark is based on a Korean language model with a retrieval-free approach . o1 model achieves the highest performance among proprietary models, while EXAONE 3.5 leads among open-sourced models. |
Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the effect of Epistemic Markers on LLM-based Evaluation (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing efforts to train large language models to generate outputs containing epistemic markers have been largely overlooked. |
| Approach: | They propose a benchmark to assess the robustness of LLM-judges to epistemic markers. |
| Outcome: | EMBER benchmarks show that LLM-judges lack robustness in presence of epistemic markers . EMber QA (2,000 instances) and IF (2823 instances) are used to evaluate outputs containing epistemological markers. |
Can You Trick the Grader? Adversarial Persuasion of LLM Judges (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly used as automated evaluators in practical settings . |
| Approach: | a study by the university of california reveals that persuasive language can bias large language models when scoring mathematical reasoning tasks. |
| Outcome: | The proposed model can bias judges when scoring mathematical reasoning tasks . Consistency causes the most severe distortion, with Consistencies leading to 8% distortion . |
LLMs can be easily Confused by Instructional Distractions (2025.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) show exceptional skill in instruction following tasks, but can become vulnerable when they are required to disregard instructions. |
| Approach: | They propose a benchmark to assess LLMs' performance under instructional distraction. |
| Outcome: | The proposed benchmark categorizes real-world instances of instructional distraction and evaluates LLMs across four instruction tasks: proofreading, rewriting, translation, and style transfer—alongside five input tasks: reasoning, code generation, mathematical reasoning, bias detection, and question answering. |
Kosmic: Korean Text Similarity Metric Reflecting Honorific Distinctions (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for text similarity measurement focus on the semantic dimension, neglecting the unique linguistic attributes found in languages like Korean. |
| Approach: | They propose a Korean text-similarity metric that encompasses the semantic and tonal facets of a given text pair. |
| Outcome: | The proposed method outperforms existing methods in Korean and other languages . it identifies which methods preserve semantics and tone while preserving similarity . |
Injecting Comparison Skills in Task-Oriented Dialogue Systems for Database Search Results Disambiguation (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing task-oriented dialogue systems handle ambiguity in database search results by randomly selecting one or few results and presenting their names to the user. |
| Approach: | They propose a task that compares properties of multiple database search results . they use a dataset to collect high-quality dialogue data and an augmented version of the SGD dataset . |
| Outcome: | The proposed task compares properties of two entities in a trade-off form based on user preferences . the proposed dataset and code will be publicized . |