Naoto Iwase
岩瀬 直人
Medical student
at Nagoya University
About
Naoto Iwase is a medical student at Nagoya University and an incoming visiting student at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI). He previously worked as a part-time engineer at Preferred Networks, Inc.
His research focuses on making large language models more reliable and useful in medicine. He develops methods and benchmarks for reliable LLM reasoning, including anytime-valid statistical inference for self-consistency, prefix-consistency evaluation for chain-of-thought reasoning, and clinical text error correction.
News
- Prefix Consistency and CITE accepted to NeurIPS 2026.
- MedRECT accepted to the ORACLE Workshop at EMNLP 2026.
- New preprints: Prefix-Denoising Consistency and OP²SD.
Publications
Accepted Papers
-
Reliable Chain-of-Thought via Prefix Consistency
TL;DR
Correct chain-of-thought traces reproduce their answer under prefix regeneration more often than wrong ones; weighting majority voting by this prefix consistency reaches plateau accuracy at up to 21× fewer tokens (median 4.6×).
[paper] [project page] [code]
-
CITE: Anytime-Valid Statistical Inference in LLM Self-Consistency
TL;DR
CITE certifies a prespecified answer as the unique mode under arbitrary data-dependent stopping, while controlling false certification without knowing the response category set in advance.
[paper]
-
MedRECT: A Medical Reasoning Benchmark for Error Correction in Clinical Texts
TL;DR
A bilingual (Japanese/English) benchmark for medical error correction built from licensing exams; across 9 LLMs, reasoning models substantially outperform standard architectures, and a fine-tuned model exceeds human expert performance.
Preprints
-
Prefix-Denoising Consistency: Test-Time Verification for Diffusion Language Models
TL;DR
Prefix-Denoising Consistency verifies diffusion language model outputs by regenerating suffixes from a fixed prefix; correct trajectories are more reproducible, improving mathematical and commonsense reasoning under matched compute.
-
Privileged Solutions or Context-Induced Teacher Behavior? Dissecting On-Policy Self-Distillation
TL;DR
Replacing a problem's paired reference with another problem-solution pair remains competitive with standard on-policy self-distillation, showing that context-induced teacher behavior matters beyond target-specific privileged information.
Awards & Honors
- Excellent Poster Presentation Award, 8th Annual Meeting of the Japanese Association for Medical Artificial Intelligence, June 2026.
- Grade Pre-1 — Pass with Distinction (S), Japan Statistical Society Certificate, January 2025.
- Student Presentation Award (Oral), 38th Annual Meeting of the Japanese Society of Computational Statistics, May 2024.
- Incoming Student Representative and Address, Nagoya University Matriculation Ceremony, April 2021 (top score among admitted applicants to the Medical Program, the university’s most selective undergraduate program).