Zexuan Wang

Research profile

Research summary

Zexuan Wang (Heilongjiang University of Chinese Medicine, China; ORCID 0000-0002-8313-5773) is an early-career, highly active researcher whose profile is name-resolved from OpenAlex bibliographic data. The record shows an h-index of 10 and i10-index of 10, with 23 works in the last three years and 45 distinct coauthors; counts by year accelerate from 1 (2016) to 11 (2026), driving an activity score of 93/100 and a collaboration score of 76/100. OpenAlex topic clusters are broad — Air Quality and Health Impacts (6), Dementia and Cognitive Impairment Research (5), Advanced Neural Network Applications (5), Toxic Organic Pollutants (3), and Gas Sensing Nanomaterials and Sensors (3) — while Evolve's catalog adds a distinct recent line: three 2026 arXiv benchmark studies of LLM agents, in all of which Wang is a middle author rather than corresponding author. HarnessDev asks whether LLMs can build and evolve their own agent harnesses, finding gains unreliable and code/search performance behind human systems; Aspire tests self-evolution from vague goals under hidden expert evaluation and reports honest but statistically thin negative results; S3Gym examines whether self-testing and self-judging translate into self-improvement, showing task-dependence. Across all three, Evolve's verdicts stress rigor and honesty about negative results alongside limited statistical power. Output, impact, focus and reproducibility dimensions remain unscored for lack of sufficient data.

Research keywords

Highlights

• HarnessDev: testing whether LLMs can build and evolve their own agent harnesses

  This 2026 arXiv benchmark finds that LLMs can construct runnable agent harnesses, but their systems trail human-built ones on code and search tasks, and evolution gains are unstable — a rigorous, candid benchmark study with medium reproducibility readiness.

• Aspire: a benchmark for self-evolution from vague goals

  The benchmark probes whether models can self-evolve when goals are vague, using a hidden expert evaluation set. Evolve's verdict: honest but statistically thin negative results — gains are rare and the evolved harness remains below the reference system; reproducibility readiness is low.

• S3Gym: when self-testing and self-judging become self-improvement

  This benchmark shows that LLM self-improvement through self-testing, self-judging and experience reuse is task-dependent. The protocol is sound and negative results are honestly reported, but tiny 3-episode evaluations lack statistical rigor; reproducibility readiness is medium.

Timeline

1. 2026HarnessDev released on arXiv, asking whether LLMs can create and evolve their own agent harnesses (middle author, not corresponding).
2. 2026Aspire posted on arXiv, benchmarking LLM self-evolution from vague goals under a hidden expert evaluation set (middle author).
3. 2026S3Gym posted on arXiv, testing whether self-testing and self-judging convert into self-improvement (middle author).
4. 2026OpenAlex records 11 works in 2026 — the highest single-year count in the profile — within 23 works over the last three years.

Papers in the Evolve catalog

• 2026-09-01

  HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?

• 2026-08-31

  Aspire: Can Models Self-Evolve from Vague Goals?

• 2026-08-31

  S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?

Frequent co-authors