Minkuk Kim

About

I am a Ph.D. student in Artificial Intelligence at Kyung Hee University, advised by Prof. Seong-Tae Kim in the AugMented Intelligence (AMI) Laboratory. My research focuses on video-language understanding, with a particular interest in long-form and streaming videos that involve extended temporal contexts.

A central theme of my work is memory-augmented video understanding: how models can retain, structure, and retrieve visual evidence over long temporal contexts. I study this problem through dense video captioning and video question answering, where models need to localize relevant events and reason over temporally distributed information.

Looking ahead, I hope to extend video understanding toward more realistic and interactive settings, such as egocentric video and embodied/VLA scenarios, where perception can eventually support action.

News

Sep 2026One paper on streaming video understanding accepted to NeurIPS 2026.
Sep 2026Started a 6-month visiting research stay at UIUC.
Jun 2026One paper on long video understanding accepted to ECCV 2026.
Dec 2024One paper accepted to AAAI 2025.
Aug 2024Completed my M.S. and started the Ph.D. program.
Feb 2024One paper accepted to CVPR 2024.

Publications

Google Scholar

TimeTraveler: a question can point to past, current or future evidence in a video stream; existing memory-based systems only look back.

TimeTraveler overall framework: the Time Dictionary is built asynchronously from the stream and the Temporal Strategy Planner routes each question to past, current or future evidence.

01 / 02

TimeTraveler: Temporal Strategy Planning with Time Dictionary for Streaming Video Understanding

Minkuk Kim, Junho Kim^*, Seong Tae Kim^*

Neural Information Processing Systems (NeurIPS), 2026

Look back, look now, or wait. The stream is kept as structured captions rather than compressed features, so long horizons lose no fine detail; an 8-bit captioner and a small retrieval budget keep every answer under a second at 1 fps.

Paper (soon)Code (soon)

ReQuest concept: uniform and similarity-based sampling miss the conclusive frame; re-thinking with question-understandable frame selection finds it.

ReQuest overall pipeline: first-thinking with sparse frames, re-thinking routing by uncertainty, then question-aware key-frame selection.

Pseudo-label generation for the frame selector from segment confidence scores.

Qualitative example: ReQuest selects the conclusive frames for a counting question.

Qualitative example: uniform and similarity sampling pick misleading frames; ReQuest picks the frames with the answer.

01 / 05

ReQuest: Rethinking-based Question-Aware Frame Selection for Long-Form Video QA

Minkuk Kim^†, Suyong Yun^†, Young Tae Kim, Jinyoung Moon, Jinwoo Choi^*, Seong Tae Kim^*

European Conference on Computer Vision (ECCV), 2026

Answer. Doubt. Look again. The model answers from its first evidence and, only when uncertainty remains, goes back for the frames the question actually needs. Less pipeline, more agent.

arXiv Code Project page

HiCM² motivation: human recall from memory versus model retrieval from a hierarchical memory.

HiCM² overall architecture with hierarchical memory read, temporal and text encoders, and a text decoder.

Hierarchical memory construction by clustering and LLM cluster summarization.

Hierarchical memory read with top-k selection and hierarchy tracking.

Qualitative example of hierarchically retrieved sentences and generated captions.

01 / 05

HiCM²: Hierarchical Compact Memory Modeling for Dense Video Captioning

Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi^*, Seong Tae Kim^*

AAAI Conference on Artificial Intelligence (AAAI), 2025

Memory in layers, the way people keep it. Captions are clustered and summarized into low, middle and high levels of abstraction, so a long video is recalled coarse-to-fine instead of searched flat. Compact, hierarchical memory of the kind long-horizon agents need, worked out on dense video captioning.

arXiv Code

CM² concept: the model queries an external memory bank and retrieves relevant sentences to caption the video.

CM² overall architecture with memory read and a versatile encoder-decoder.

Memory read: retrieval with temporal anchors and feature aggregation.

Versatile encoder and decoder with visual and textual cross-attention.

Qualitative example of retrieved sentences and generated captions.

01 / 05

Do You Remember? Dense Video Captioning with Cross-Modal Memory Retrieval

Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi^*, Seong Tae Kim^*

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

Recall before you describe. An external memory of sentences is retrieved against what the model sees, and the recalled cues guide both where an event is and how to say it. Retrieval-augmented generation, for video: where this line of work began.

arXiv Code

^* Corresponding author · ^† Equal contribution