Explore the research
We study generative AI for images, text, and multimodal data, including data augmentation, curriculum learning, active learning, learning support, and validation of generated content.
The research emphasizes both generation quality and how generated outputs can be evaluated and used responsibly in downstream tasks.
We also study LLM-as-a-Judge systems as evaluators and measurement instruments, focusing on the stability, bias, and criterion shifts of the evaluator itself.
Why It Is Difficult
Evaluation of generative models includes subjective components, making it difficult to design stable quantitative metrics.
Quality, diversity, learning stability, semantic validity, bias, and downstream utility often trade off with one another.
Approach
We design multi-perspective evaluation protocols, analyze correlation with human judgment, and introduce curriculum learning, active learning, and verification loops to improve learning efficiency and output reliability.
For LLM judges, we separately evaluate output validity, non-tie preferences between identical answers (dark current), and selection of a specified target under a declared grading criterion.
Evaluation
We combine quantitative indicators such as FID, IS, CLIP-based metrics, and task-specific downstream scores with human evaluation and semantic consistency checks.
Current Questions
- 1Evaluation design that balances image quality and diversity
- 2Active learning and generative support for efficient annotation
- 3Verification of fairness, bias, and semantic validity in generated outputs
- 4Reliability analysis and measurement protocols for LLM-as-a-Judge systems
Related publications (6)
Dynamic BGP Route Control Based on Real-Time Performance Information Using an LLM with a Verification Mechanism
Combines LLM-based decisions with a verification mechanism to study network route control using real-time performance information.
Tsunogai, Kento, Yamasaki, Shugo, Kawabe, Kaito, Usami, Hiroyasu
Proceedings of the 2026 Tokai-Section Joint Conference on Electrical, Electronics, Information, and Related Engineering, 2026 (8)
2B-2-7, p. 154, Network Systems, August 30, 2026
A Stabilization Method for Automatic Educational Question Generation Based on Self-consistency in Chain-of-Thought Reasoning
Studies automatic generation of educational questions using self-consistency in chain-of-thought reasoning.
Shimotsuji, Nayuta, Usami, Hiroyasu, Nakane, Jin
23rd Workshop on Informatics (WiNF 2025), 2025
P1A-3
Verification of Semantic Biosynthesis in Image Generative AI Through a Modality Conversion-based Generation and Verification Cycle
Studies semantic correspondence in generated images through a generation-and-verification cycle across modalities.
Izuhara, Keima, Tominari, Hiroki, Usami, Hiroyasu
23rd Workshop on Informatics (WiNF 2025), 2025
P1A-30
A Study of Purpose-oriented Prompts for Image Generative AI
Studies how to design image-generation prompts around the intended objective.
Fukuoka, Reiya, Tominari, Hiroki, Usami, Hiroyasu, Iwahori, Yuji
Proceedings of the 55th Tokai Fuzzy Workshop, Japan Society for Fuzzy Theory and Intelligent Informatics, 2024
P1-07
Research and Evaluation of Effective Prompts for Stable Diffusion-based AI Image Generation Tools
Studies prompt design and evaluation for Stable Diffusion-based image generation.
Tominari, Hiroki, Fukuoka, Reiya, Usami, Hiroyasu
21st Workshop on Informatics (WiNF 2023), 2023
S3C-4
A Text-to-image Model Considering Latent Information Specific to Japanese
Studies text-to-image modeling that considers information specific to the Japanese language.
Fukuoka, Reiya, Yamazaki, Daiki, Yamane, Haruki, Usami, Hiroyasu, Suzuki, Yuri
20th Workshop on Informatics (WiNF 2022), 2022
S-4A-7