Shunichiro Hayashi

Research focus

Code Generation & Program Synthesis (2)Distributed Systems & Hardware (1)Architecture Design (Transformers, SSMs, MoE) (1)Inference & Quantization (1)

Frequent co-authors

Daichi Mukunoki (2)Tetsuya Hoshino (2)Takahiro Katagiri (2)Ryo Mikasa (1)

Papers (2)

Jun 4, 2026

Affiliation:Jun 4, 2026·also Nagoya

LLM-Based Porting of Optimized C++ to CUDA Through Deoptimization and Reoptimization

Deopt-Reopt can significantly accelerate CUDA translation for certain HPC kernels, but its advantages are kernel-dependent and not universally applicable.

Daichi Mukunoki, Ryo Mikasa, Shunichiro Hayashi +2

Code Generation & Program Synthesis Distributed Systems & Hardware

Apr 19, 2026

Apr 19, 2026·also Affiliation:

Layer-wise MoE Routing Locality under Shared-Prefix Code Generation: Token-Identity Decomposition and Compile-Equivalent Fork Redundancy

LLMs generating code from a shared prefix exhibit surprisingly high MoE routing similarity even when producing different tokens, but this similarity varies significantly across layers, revealing potential optimization points.

Shunichiro Hayashi, Daichi Mukunoki, Tetsuya Hoshino +1

Architecture Design (Transformers, SSMs, MoE)Code Generation & Program Synthesis Inference & Quantization

Search

Shunichiro Hayashi

Research focus

Frequent co-authors

Papers (2)