CV

Research Intern

NVIDIA Research

Working on NVIDIA’s Omni models on the speech side — agentic ASR, speech translation, TTS, and evaluation.

• Developed a novel agent-based intelligence-quotient evaluation for speech LLMs and cascaded pipelines that outperforms existing metrics in aligning with human preference (SpeechIQ, ACL 2025 Main).
• Introduced a voice-agentic framework that teaches an omni-model when to trust itself versus when to consult external audio perception (Speech-Hands, ACL 2026 Oral).

Research Assistant

LLM-jp

Japanese instruction data via GPT-4 self-instruction and an automatic evaluation benchmark for Japanese LLMs.

• Co-led the ja-vicuna-qa-benchmark project.
• Co-first-authored LREC-COLING 2024 paper on rapid Japanese instruction data + benchmark construction.

Research Assistant

SIP — Cross-ministerial Strategic Innovation Promotion Program

Continually pre-trained general Japanese LLMs on Japanese medical domain corpus and instruction-tuned on Japanese medical tasks.

• Built the training framework + open tutorial for Megatron-LM parallel training on MDX Cloud.
• Two EMNLP 2025 papers (Main + Findings) on Japanese medical LLMs.

Research Fellow (DC2)

Japan Society for the Promotion of Science (JSPS)

DC2 fellowship on Multi-document and Multi-modal Information Extraction (acceptance rate 19.1%).

Research Intern

Alibaba DAMO Academy

Data selection and training curriculum for continually pre-training and instruction-tuning Chinese Legal LLMs.

• Built a multi-step retrieval-augmented system for robust, updated legal QA.
• 1st-authored ACL 2024 Findings paper on Adapt-Retrieve-Revise for Chinese Legal LLMs.

PhD Candidate

Kyoto University

Informatics (Intelligence Science & Technology)

• Advised by Prof. Sadao Kurohashi and Prof. Chenhui Chu.
• JSPS Research Fellowship for Young Scientists (DC2).

MS

Kyoto University

Informatics (Intelligence Science & Technology)

B.E.

Zhejiang University

Energy Engineering

ACL 2025 Best Industrial Paper Honor Mention

For NeKo (4th author) — task-guided mixture-of-experts for cross-modality post-recognition error correction.

JSPS Research Fellowship for Young Scientists (DC2)

Acceptance rate 19.1%. Multi-document and Multi-modal Information Extraction.

University Fellowship in Informatics, Kyoto University

3rd Prize, College Industrial Design Competition