Yuma Ichikawa

Research profile

Research summary

Yuma Ichikawa is a researcher at Fuji Electric in Japan whose work spans machine-learning methods for generative modeling and, more recently, efficient large language model (LLM) inference. OpenAlex records show 23 works since 2022, 19 of them in the last three years, with output accelerating to 10 works in 2026 — a peak year that includes the two arXiv papers Evolve has reviewed. Topic modeling places his core contributions in generative adversarial networks and image synthesis (7 works), model reduction with neural networks (5), and stochastic gradient optimization (4), alongside smaller clusters in neural network applications (3) and quantum computing algorithms (3). The 2026 papers mark a pivot toward LLM inference and test-time compute: CASA replaces scalar sensitivity with Hessian-based costs and cross-layer search for mixed-precision LLM quantization, and VRR reuses verifier feedback to generate better candidates. On both, Ichikawa is the last-listed author rather than the corresponding author. Evolve's verdicts note strong low-bit quantization gains but thin baselines and statistics for CASA, and modest, significance-untested gains for VRR; both carry medium reproducibility readiness. Collaboration metrics are narrow — 2 distinct coauthors recorded, chiefly Akiyoshi Tomihari and Akihiro Yoshida — and citation-based output and impact scores are not yet computable, making this an early-stage, industrially anchored profile.

Research keywords

Highlights

• Hessian-aware mixed-precision quantization (CASA)

  On the 2026 arXiv paper 'Beyond Scalar Sensitivity: Activation-Aware Mixed-Precision LLM Quantization with Cross-Layer Refinement', Evolve notes that CASA replaces scalar sensitivity with Hessian-based costs and a cross-layer search over bit allocation, delivering strong gains at low bit-widths — though baselines and statistical comparisons remain thin. Ichikawa is the last-listed author.

• Verifier feedback reused for candidate improvement (VRR)

  In 'LLM-as-an-Improver: Turning Verification into Better Candidates' (2026), the VRR method reuses LLM verifier feedback to repair and regenerate candidates, including honest recovery from an all-wrong pool; Evolve judges the gains modest and not yet tested for statistical significance. Ichikawa again appears as last author.

• From GANs to LLM inference: a 2026 acceleration

  Annual publication counts rose from 1 work in 2022 to 10 in 2026, with 19 works in the last three years; the 2026 output pivots from a generative-modeling and optimization base toward LLM quantization and test-time compute, both reviewed by Evolve at medium reproducibility readiness.

Timeline

1. 2022First recorded OpenAlex work; the start of a publication record that totals 23 works by 2026.
2. 20233 works recorded — output triples from the 2022 baseline.
3. 20245 works recorded, the highest annual count to that point.
4. 20254 works recorded, holding near the 2024 level.
5. 2026Peak year: 10 works recorded, including the two arXiv papers reviewed by Evolve — CASA on mixed-precision LLM quantization and VRR on verifier-feedback candidate improvement.

Papers in the Evolve catalog

• 2026-09-22

  Beyond Scalar Sensitivity: Activation-Aware Mixed-Precision LLM Quantization with Cross-Layer Refinement

• 2026-09-18

  LLM-as-an-Improver: Turning Verification into Better Candidates

Frequent co-authors