Jiahao Zhang

Research Scientist @ SB Intuitions · Ph.D., The University of Osaka

I am a Research Scientist at SB Intuitions, working on research and development for Responsible AI. I received my Ph.D. from the D3 Center (formerly the Institute for Datability Science), The University of Osaka, where I was advised by Prof. Yuta Nakashima and Prof. Hajime Nagahara, and my M.S. from the Graduate School of Medicine, Osaka University.

My research interests include Responsible AI, visual in-context learning, and the safety and alignment of vision–language models. I am also interested in medical image analysis and AI applications for healthcare.

Portrait of Jiahao Zhang

News

Selected Recent Work

A closer look at four recent projects. * denotes (co-)first author.

Safety-neuron analysis overview
EMNLP 2026 (Main)

VLM Safety Interpretability

Do VLMs Share Safety Neurons Across Modalities?

Jiaxuan Li*, Jiahao Zhang*, Duc Minh Vo, Huy H. Nguyen, Pride Kavumba, Koki Wataoka

Vision–language models often comply with a harmful request once it arrives as an image, even when their LLM backbone would refuse the very same words in text. This work traces that gap causally at the neuron level across 10 VLMs, using a two-stage detection pipeline whose iterative ablation accounts for self-repair, plus two modality-isolated benchmarks, ViSafe-Detect and ViSafe-Eval. Text safety turns out to be sharply localizable — roughly 88 neurons, under 0.01% — and is the dominant refusal pathway, while visual safety stays diffuse: text safety concentrates in about 5 subspace directions where visual safety needs 50 or more.

arXiv PDF Project

PANICL overview
arXiv

Visual In-Context Learning Training-free

PANICL: Mitigating Over-Reliance on Single Prompt in Visual In-Context Learning

Jiahao Zhang*, Bowen Wang, Hong Liu, Yuta Nakashima, Hajime Nagahara

Visual in-context learning conditions on a single input–output image pair, which makes predictions biased and unstable. PANICL is a training-free framework that aggregates multiple in-context pairs instead, smoothing assignment scores across them. It improves foreground segmentation, single-object detection, colorization, multi-object segmentation and keypoint detection over strong baselines, and holds up under both dataset-level and label-space domain shifts.

arXiv PDF

InMeMo overview
WACV 2024

Visual In-Context Learning Prompt Learning

Instruct Me More! Random Prompting for Visual In-Context Learning

Jiahao Zhang*, Bowen Wang, Liangzhi Li, Yuta Nakashima, Hajime Nagahara

Visual in-context learning hands a large frozen model a single input–output image pair to exemplify the task, so accuracy hinges on how good that pair is. InMeMo augments the in-context pair itself with a learnable perturbation, letting the demonstration adapt while the model stays frozen. This lightweight training lifts mIoU by 7.35 on foreground segmentation and 15.13 on single-object detection over the no-learnable-prompt baseline.

arXiv PDF CVF Code

DiReCT pipeline
NeurIPS 2024

LLM Reasoning Benchmark

DiReCT: Diagnostic Reasoning for Clinical Notes via Large Language Models

Bowen Wang, Jiuyang Chang, Yiming Qian, Guoxin Chen, Junhao Chen, Zhouqiang Jiang, Jiahao Zhang, Yuta Nakashima, Hajime Nagahara

LLMs answer medical exam questions well, but rarely show reasoning a clinician can audit. DiReCT is a benchmark of 511 clinical notes, each annotated by physicians with the full diagnostic path from observation to final diagnosis, paired with a diagnostic knowledge graph. Evaluating leading LLMs on it exposes a wide gap between their reasoning and that of human doctors.

arXiv PDF Code

Publications

Selected works; see Google Scholar for the full list. * denotes (co-)first author.

International Conferences

Do VLMs Share Safety Neurons Across Modalities?

Jiaxuan Li*, Jiahao Zhang*, Duc Minh Vo, Huy H. Nguyen, Pride Kavumba, Koki Wataoka

EMNLP 2026 (Main) arXiv PDF Project

DiReCT: Diagnostic Reasoning for Clinical Notes via Large Language Models

Bowen Wang, Jiuyang Chang, Yiming Qian, Guoxin Chen, Junhao Chen, Zhouqiang Jiang, Jiahao Zhang, Yuta Nakashima, Hajime Nagahara

NeurIPS 2024 arXiv PDF Code

Instruct Me More! Random Prompting for Visual In-Context Learning

Jiahao Zhang*, Bowen Wang, Liangzhi Li, Yuta Nakashima, Hajime Nagahara

WACV 2024 arXiv PDF CVF Code

Journal Articles

ILD-Slider: A Parameter-Efficient Model for Identifying Progressive Fibrosing Interstitial Lung Disease from Chest CT Slices

Jiahao Zhang*, Shoya Wada, Kento Sugimoto, Takayuki Niitsu, Kiyoharu Fukushima, Hiroshi Kida, Bowen Wang, Shozo Konishi, Katsuki Okada, Yuta Nakashima, Toshihiro Takeda

J. Imaging Paper PDF

E-InMeMo: Enhanced Prompting for Visual In-Context Learning

Jiahao Zhang*, Bowen Wang, Hong Liu, Liangzhi Li, Yuta Nakashima, Hajime Nagahara

J. Imaging Paper PDF arXiv Code

Automatic Creation of Annotations for Chest Radiographs Based on the Positional Information Extracted from Radiographic Image Reports

Bowen Wang, Toshihiro Takeda, Kento Sugimoto, Jiahao Zhang, Shoya Wada, Shozo Konishi, Shirou Manabe, Katsuki Okada, Yasushi Matsumura

CMPB Paper

Preprints

PANICL: Mitigating Over-Reliance on Single Prompt in Visual In-Context Learning

Jiahao Zhang*, Bowen Wang, Hong Liu, Yuta Nakashima, Hajime Nagahara

arXiv PDF

VASCAR: Content-Aware Layout Generation via Visual-Aware Self-Correction

Jiahao Zhang*, Ryota Yoshihashi, Shunsuke Kitada, Atsuki Osanai, Yuta Nakashima

arXiv PDF

Explainable Image Recognition via Enhanced Slot-Attention Based Classifier

Bowen Wang, Liangzhi Li, Jiahao Zhang, Yuta Nakashima, Hajime Nagahara

arXiv PDF

Experience

Education

Awards

Academic Activities