KUDoc

HiKEY introduces a hierarchical multimodal retrieval framework for open-domain document question answering, combining global routing, local ranking, and structured evidence assembly across long and visually complex documents.

Proposes an entity-aware sub-table generation framework for question answering over real-world multi-table settings, addressing challenges of entity disambiguation and cross-table reasoning in structured document understanding.

Introduces a dependency-aware chunking framework that leverages LVLMs to capture structural and semantic dependencies across multiple pages and documents for improved retrieval.

Proposes a hierarchical multimodal chunking pipeline that integrates text, layout, and visual signals to boost RAG performance on long, complex industrial documents.

Introduces a hierarchical taxonomy of OCR errors and a synthetic data contamination strategy that injects realistic OCR-like noise at the character, word, and structural level — enabling robust document reconstruction without costly real-error annotations.

Presents a domain-specific RAG system for power plant maintenance QA, combining StyleDFS retrieval and instruction-tuned generation for industrial deployment.

초거대 언어 모델을 활용하여 교육 도메인에서 구조화된 정보를 추출하는 과업을 제안한다. 교과서 단락 텍스트로부터 핵심어를 계층적으로 추출하는 평가 데이터셋을 구축하고, 오픈 소스 및 사유 LLM에 대한 In-Context Learning 방식의 평가를 진행한다.

Proposes a post-hoc language identity removal approach using sparse autoencoder editing to improve cross-lingual retrieval performance in multilingual language models.

Investigates which parameters within LLMs are responsible for coding capabilities through mechanistic analysis, offering insights for targeted fine-tuning and model understanding.

Introduces a multilingual and multimodal alignment benchmark for evaluating cultural grounding in vision-language models across diverse linguistic and cultural contexts.

Proposes a mechanistic diagnostic framework for LLM benchmark evaluation, identifying systematic failure modes and biases in standard benchmarks through profiling techniques.

Introduces a continuous retrieval framework for on-the-fly Korean legal knowledge editing, enabling dynamic updates to LLM knowledge bases in domain-specific legal contexts.

Leverages syntactic retrieval with LLMs to improve automatic term extraction, achieving stronger precision and recall on domain-specific terminology identification tasks.