Fenghua Yang

Research profile

Research summary

Fenghua Yang is a researcher at the University of Michigan whose work sits at the interface of statistics, operations research, and machine learning, with a distinctive signature: putting large language models inside rigorously validated statistical and decision-making pipelines. Evolve has reviewed two 2026 arXiv papers. The first, first-authored, presents automated design of inventory policy with LLMs: an LLM-plus-solver evolutionary search over policy classes that discovers interpretable lost-sales policies beating optimized base-stock policies by about 30% on average, with strong ablations and high reproducibility readiness. The second, last-authored, proposes SCALE, a cost-aware framework for human-AI-powered hypothesis testing that selectively uses AI scoring (LLM-as-judge) and sequentially escalates to human raters; it carries finite-sample validity and first-order optimality guarantees, though evidence is limited to synthetic simulations. OpenAlex records 8 works in the past three years — all dated 2026 — and 10 distinct coauthors, with a frequent-coauthor circle including Stefanus Jasin, Dae Woong Ham, Xuejun Zhao, and Parshan Pakiman. Topic signals cluster around game theory and applications, machine learning and algorithms, advanced bandit algorithms, simulation techniques, and statistical methods in clinical trials. Because the record is so recent, citation-based output and impact metrics are not yet available, making this an early-stage bibliometric profile with strong methodological signals.

Research keywords

Highlights

• LLM-driven evolutionary search designs inventory policies that beat base-stock by 30%

  As first author, Yang's exploratory study couples large language models with a solver-in-the-loop evolutionary search over policy classes, yielding interpretable lost-sales policies that outperform optimized base-stock policies by about 30% on average, with strong ablations; Evolve rates its reproducibility readiness high while noting thin benchmarks.

• SCALE: cost-aware human-AI hypothesis testing with finite-sample guarantees

  As last author, Yang contributes a framework for AI-scored hypothesis testing that selectively escalates to human raters under a sequential design; Evolve's verdict highlights cost-minimal AI+human testing that is finite-sample valid and first-order optimal, with rigorous theory but evidence limited to synthetic simulations (medium reproducibility readiness).

• A two-front agenda: LLMs as both scorer and search engine

  Across the two reviewed papers, LLMs play opposite roles — a cost-aware scoring engine inside statistically valid sequential testing, and a heuristic search operator for designing operations policies — evenly spanning human–AI statistical inference and LLM-based inventory management (a 50/50 field distribution).

Timeline

1. 2026First-authored 'Automated Design of Inventory Policy with Large Language Models: An Exploratory Study' posted to arXiv.
2. 2026Last-authored 'Human-AI-Powered Hypothesis Testing: Cost-Aware Selective AI Scoring and Sequential Human Escalation' posted to arXiv.
3. 2026OpenAlex records 8 works in this single year with 10 distinct coauthors — the entire visible publication record to date.

Papers in the Evolve catalog

• 2026-09-24

  Human-AI-Powered Hypothesis Testing: Cost-Aware Selective AI Scoring and Sequential Human Escalation

• 2026-09-08first author

  Automated Design of Inventory Policy with Large Language Models: An Exploratory Study

Frequent co-authors