Presentation Details

10:00~10:30  [개회사, 환영사, 성대 AI성과 소개]

• 이지형 교수

10:30~11:45 [교수 연구 발표]

• 발표교수: 심규홍 교수

• 발표제목: Efficient Inference for Long-Context Multimodal LLMs

• 초록: As multimodal large language models (MLLMs) continue to extend their context length, a single model can now integrate information from text, audio, video, and embodied signals. Despite this progress, deploying ultra-long-context models in real systems remains difficult because of practical memory constraints and strict latency requirements. In this talk, I will outline recent approaches designed to address these challenges, with particular attention to techniques that compress the key–value (KV) cache. I will close by highlighting open research directions and practical considerations for building scalable and efficient multimodal LLM inference pipelines.

• 발표교수: 이성길 교수

• 발표제목: DC4GS: Directional Consistency-Driven Adaptive Density Control for 3D Gaussian Splatting

• 초록: We present a Directional Consistency (DC)-driven Adaptive Density Control (ADC) for 3D Gaussian Splatting (DC4GS). Whereas the conventional ADC bases its primitive splitting on the magnitudes of positional gradients, we further incorporate the DC of the gradients into ADC, and realize it through the angular coherence of the gradients. Our DC better captures local structural complexities in ADC, avoiding redundant splitting. When splitting is required, we again utilize the DC to define optimal split positions so that sub-primitives best align with the local structures than the conventional random placement. As a consequence, our DC4GS greatly reduces the number of primitives (up to 30% in our experiments) than the existing ADC, and also enhances reconstruction fidelity greatly.

• 발표교수: 이선재 교수

• 발표제목: GUI Agent: Practical Solutions to automating all digital tasks

• 초록:  Graphical User Interface (GUI) Agents (i.e., Computer Use Agent) represent a new paradigm in automation, where AI systems directly interact with existing software interfaces to perform tasks traditionally executed by humans. This talk provides a practical overview of GUI Agents as an emerging foundation for end-to-end task automation across digital environments. I will introduce the core concept of GUI Agents, and explain the key technical components that enable them. I will also highlight recent research trends such as Computer-use foundation models and Neuro-hybrid architecture. Finally, I will briefly present ongoing research from our lab, focusing on benchmarking, reliability, and verification of GUI Agents for large-scale, real-world automation.

13:00~14:15 우수논문 포스터 세션 #A

• Poster Session Details 참조 

14:20~15:20 [Student Spolight]

• Student Spolight 1

• 유민종 석박사통합과정

• 발표제목: Skill-centric Learning Framework for Open-domain Embodied AI

• 초록: Embodied AI systems in open-domain environments face three fundamental challenges: handling diverse tasks across multiple domains, adapting to dynamic environmental changes, and transferring policies across different physical embodiments. We present a unified skill-centric learning framework. Our approach leverages skills, semantically meaningful action sequences, as a core abstraction, decomposing decision-making into a high-level task planner for skill composition and a low-level controller for embodiment-specific action reconstruction. We propose novel methods for each challenge: retrieval-augmented planning with temporal knowledge graphs for dynamic adaptation, and world model implanting with skill disentanglement for cross-embodiment transfer. Comprehensive experiments across robotic manipulation, household tasks, and autonomous driving consistently outperform state-of-the-art methods, establishing a foundation for robust embodied agents in real-world deployment.

• Student Spolight 2

• 이주찬 석박사통합과정

• 발표제목: Efficient Neural Fields for Visual Signal Representation

• 초록:  Neural fields have become a powerful framework for representing signals such as images, videos, and 3D scenes in a continuous domain. However, their practical use is often limited by trade-offs among accuracy, compactness, and speed. In this presentation, I will begin with a brief overview of neural fields, then introduce more efficient architectures that better navigate these trade-offs. I will then focus on task-driven applications in video and 3D reconstruction, discussing how to design compact, fast, and accurate neural field representations that are practical for real-world applications.

• Student Spolight 3

• 정희수 박사과정

• 발표제목: Rethinking Graph Self-Supervised Learning: Advances, Analyses, and Explanations

• 초록: This work examines graph self-supervised learning (SSL) by integrating three complementary advances in methodology, theoretical understanding, and explainability. CIMAGE refines masked graph auto-encoding through a conditional–independence–guided masking strategy that reduces redundancy and enhances the relevance of reconstructed signals, yielding linearly separable and robust representations. BSG provides an information-theoretic decomposition of graph SSL objectives, analyzes the role of embedding smoothness in shaping model performance, and introduces a balanced loss that reconciles competing smoothness effects across diverse tasks. HINT-G extends the scope of GNN explanation for SSL by leveraging influence functions to identify both existing and non-existent edges that significantly shape learned representations, enabling task-agnostic explanations. Together, these contributions offer a unified perspective on the mechanisms, trade-offs, and explanatory principles of graph self-supervised learning.

• Student Spolight 4

• 이진섭 석박사통합과정

• 발표제목: Robust and Generalized Learning in Real-world Scenarios

• 초록: Real-world data often include noisy labels, domain variation, and structural complexity, creating practical challenges for developing reliable and generalizable models. In this talk, I present four scenarios related to real-world data environments and applications. First, LSL incorporates structural information from data distribution to prevent overfitting to incorrect labels. Second, DomCLP enhances generalization to unseen domains by combining domain-aware contrastive learning with prototype mixup to learn robust and domain-invariant features. Third, DCG-SQL proposes a deep contextual schema link graph that jointly represents questions and database schemas, enabling more accurate demonstration retrieval in Text-to-SQL. Lastly, RA-RFT enhances relevance discrimination and fine-grained semantic reasoning, which allows it to effectively refuse hard-irrelevant queries in Video Temporal Grounding.

15:30~16:45 [교수 연구 발표]

• 발표교수: 우홍욱 교수

• 발표제목: Pathways to AGI: Agentic, Embodied, and Physical AI 

• 초록: How do we transition from language models that predict text to intelligent agents that understand the world? At the CSI-Agent Lab, we believe the path to Artificial General Intelligence (AGI) lies in systems that, like humans, learn continuously through interaction and embodiment. In this talk, I will present a unified framework for AGI that triangulates agentic AI (multi-agent synergy), embodied AI (experiential learning), and physical AI (world understanding). By showing recent projects from our lab, I will demonstrate how we are moving away from isolated capabilities and toward a holistic architecture where agents not only process information but interact with, learn from, and physically reason about their environments.

• 발표교수: 민동문 교수

• 발표제목: Quantum Machine Learning: System Architect’s Perspective

• 초록: This seminar introduces the promising advantages of Quantum Machine Learning (QML) while analyzing the critical barriers preventing its practical realization. I will then briefly overview the system-level innovations and architectural efforts currently underway to overcome these challenges.

• 발표교수: 최윤석 교수

• 발표제목: Large Language Models for Code Intelligence

• 초록: This talk provides an overview of the challenges LLMs must address for Code Intelligence, including structural understanding, robustness, hallucination, and SQL/Table reasoning, and discusses how to better handle programming languages.

16:45~18:00  [우수논문 포스터 세션 #B]

• Poster Session Details 참조 

18:00     [폐회]