Drag the paper, or use arrow keys to move it. Press Enter to turn it over, and Home to reset its position.

AI RESEARCHER · SEOUL

A WORLD OF POSSIBILITIES

A curious mind. An open world.

Inha Kang.

Making AI see, reason
& imagine what’s next.

Vision–language models. Physical intelligence. Agentic systems.
I turn model failures into new capabilities.

KAIST AIPh.D. studentKRAFTONResearch intern
Ideas in motion

A small universe of things I’m curious about.

Qualcomm Innovation Fellowship Korea

Selected as a QIFK 2026 Finalist.

FINALIST

NeurIPS 2026

Recognized as a Top Reviewer for main-track reviewing.

TOP REVIEWER

A new chapter at KRAFTON

Joined as a research intern, working on LLMs and agentic tool use.

↗

01 / SELECTED RESEARCH

A closer look.
A better model.

From understanding what a model misses
to changing what it can do.

All 10 publications
ECCV 2026

What do models
really see?

Research overview: Blind to Position, Biased in Language: Probing Mid-Layer Representational Bias in Vision-Language Encoders for Zero-Shot Language-Grounded Spatial Understanding
TRAINING-FREESpatial grounding from mid-layer representations
01

Blind to Position, Biased in Language: Probing Mid-Layer Representational Bias in Vision-Language Encoders for Zero-Shot Language-Grounded Spatial Understanding

Co-first author

Na Min An*, Inha Kang*, Minhyun Lee, Hyunjung Shim

Read paper
ICLR 2026

Understanding
the word “not”.

Research overview: What “Not” to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging
+10.8NMS-AP points · up to, on OVDEval negation
02

What “Not” to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging

First author

Inha Kang, Youngsun Lim, Seonho Lee, Jiho Choi, Junsuk Choe, Hyunjung Shim

Read paper
CVPR 2026

Better forecasts.
Fewer false alarms.

Research overview: Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization
−47.3%False-alarm rate · compared with SFT alone
03

Real-Time Long Horizon Air Quality Forecasting via Group-Relative Policy Optimization

First author

Inha Kang, Eunki Kim, Wonjeong Ryu, Jaeyo Shin, Seungjun Yu, Yoon-Hee Kang, Seongeun Jeong, Eunhye Kim, Soontae Kim, Hyunjung Shim

Read paper
EMNLP 2025 · Findings

A sense of
three dimensions.

Research overview: 3D-Aware Vision-Language Models Fine-Tuning with Geometric Distillation
2D → 3DGeometric knowledge, distilled into VLMs
04

3D-Aware Vision-Language Models Fine-Tuning with Geometric Distillation

Co-author

Seonho Lee*, Jiho Choi*, Inha Kang, Jiwook Kim, Junsung Park, Hyunjung Shim

Read paper
There’s more where that came from.Explore the full research archive
A WINDOW INTO THE PHYSICAL WORLD
EAST ASIA · AIR QUALITY · OPEN DATA
DATASET AVAILABLE

A little further
into the atmosphere.

Meet PLUME: direct multiday particulate forecasting with physically conditioned correction and ordered uncertainty over East Asia.

The public dataset brings together CMAQ reanalysis, gridded ground-station observations and trained model weights for studying air-quality forecasting.

2016–2024
hourly record
0.25°
shared spatial grid
+120 h
forecast horizon
Explore the dataset

Dataset released · Paper archive forthcoming

THE THREAD THROUGH MY WORK

Understand the failure.
Build the next capability.

01 / SEE & UNDERSTAND↘
01

See &
understand.

I probe how vision-language models represent space, language and negation, then use those insights to improve grounding and reasoning.

  • Vision–language models
  • Explainable AI
  • Representation learning
02

Predict the
physical world.

I connect foundation models with physical dynamics, uncertainty and decision-aware objectives. My current work explores task-specific world models for multi-day forecasting.

  • World models
  • Physical understanding
  • Reinforcement learning
03

Turn reasoning
into action.

At KRAFTON, I work on reliable agentic tool use: separating model errors from evaluation artifacts, curating training data and designing verifiable post-training tasks.

  • LLM foundation models
  • Agentic tool calling
  • Data curation & evaluation

02 / THE PERSON BEHIND THE PAPERS

Research depth.
Builder energy.

Across disciplines.
Close to the real world.

I’m Inha, a researcher who likes to understand why things fail—and build something better from there.

I’m a Ph.D. student at KAIST AI, advised by Prof. Hyunjung Shim in the Computer Vision and Machine Learning Lab, and a research intern at KRAFTON.

My path runs through 3D medical imaging, multimodal perception, vision-language reasoning and now agentic systems. Before my Ph.D., I spent two years at FaindersAI, building data pipelines and perception systems for real-world retail environments.

That mix shapes my approach: careful model analysis, better data, and research that carries through to working systems.

Where I’ve been.
What I’m building.

RESEARCH × ENGINEERING

EXPERIENCE

AUG 2026 — PRESENT

KRAFTON

NOW

Research Intern · LLMs & Agentic Systems

Benchmark and failure analysis, role-aware tool-use data curation, and executable RL tasks with state-based verification.

2022 — 2024

FaindersAI

Deep Learning Researcher

VLM-based segmentation-data curation, continual learning, heterogeneous sensor fusion and real-time retail change detection.

Earlier research experience

2019 — 2020

Machine Learning and Vision Lab

Research Intern · Long-tailed object detection for autonomous driving.

2017 — 2018

Aerospace Propulsion Lab

Research Intern · Parallel computing and GPU acceleration.

EDUCATION

2024 — 2028 EXPECTED

KAIST Ph.D. · Artificial Intelligence

CVML Lab · Prof. Hyunjung Shim · GPA 4.3/4.3

2020 — 2022

KAIST M.S. · School of Computing

CGV Lab · Prof. Jinah Park

2016 — 2020

GIST B.S. · EECS

Salutatorian · Best Thesis Award
Minors in Business Administration and Sociology

2017

UC Berkeley Visiting Study

Computer Science

03 / A FEW MILESTONES

Work that
gets noticed.

Recognition for research, reproducibility
and contributions to the community.

2026 · QUALCOMM AI RESEARCH

QIFK
Finalist.

Qualcomm Innovation Fellowship Korea

2026 · NEURIPS

Top
Reviewer.

Main-track reviewing service

2026 · CVPR

Compute
Transparency.

Compute Transparency Champion Award

2021 + 2022 · MICCAI MOOD

First place.
Twice.

2021 challenge · 2022 sample-level task

Also along the way Best Presentation Award, KTCP 2022 · Best Thesis Award, GIST 2020

Talk & patent Invited speaker, KOSMII 2022 · Data Collection Apparatus and Method, Korean patent publication No. 10-2025-0031529

CURIOSITY DOESN’T STAY IN ONE LANE

Beyond the paper.

Immersive interfaces, virtual fashion, mobile AI, and the joy of building things.

Play & build