Yuhao Wu — LLM architectures and agent-evaluation researcher at CUHK-Shenzhen; first author of the HarnessDev benchmark
3
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Output is not yet scored: the OpenAlex record shows h-index 0 and i10-index 0 with only 3 works in the recent three-year window, too little to measure cumulative publication volume. This axis would track total works and their year-by-year trajectory.
Impact—
Impact is not yet scored: no citation counts or citation-by-year data are available in the source record. This axis would measure how often the researcher's papers are cited and their influence within and beyond their fields.
Focus0/100
The focus score is 0, consistent with an even 50/50 split in the field distribution between machine learning on large language model architectures and LLM agents / AI systems evaluation in software engineering. Topic counts — topic modeling (3), NLP techniques (2), and three single-count topics — reinforce breadth across subfields rather than depth in one.
Collaboration68/100
Collaboration scores 68. The record lists 29 distinct coauthors against only 3 works in the recent three-year window — a wide per-paper network — with 12 recurring collaborators including Shaowen Wang, Ge Zhang, and Jiaheng Liu. Authorship spans first author on HarnessDev to middle positions on SMELT and RePro.
Activity40/100
Activity scores 40. Counts by year show 3 works, all falling in 2026, matching the 3 works in the recent three-year window. Output is therefore recent and front-loaded rather than spread over multiple years, consistent with an early-stage record whose indexed history is short.
Reproducibility38/100
Reproducibility scores 38 as the mean across 2 reviewed papers. HarnessDev is rated medium: a benchmark study that openly reports where LLM-built harnesses lose to human systems on code and search. SMELT is rated low: its compute-savings claims rest on single runs with proprietary, unverifiable training. Practice is real but uneven.
Research summary
Yuhao Wu, affiliated with the Chinese University of Hong Kong, Shenzhen (CN), works across two equally weighted lines of large language model research: LLM architectures — mixture-of-experts looped transformers and scaling laws — and LLM agent evaluation in software engineering, each accounting for half of the indexed field distribution. This profile is name-resolved from public OpenAlex bibliographic data; h- and i10-index values in the record read zero, so no output or impact score is claimed. Three 2026 arXiv papers anchor the catalog. HarnessDev (first author) asks whether LLMs can create and evolve their own agent harnesses; Evolve's review finds LLMs can build runnable harnesses but lag human-built systems on code and search tasks, with unstable evolution gains — a rigorous, candid benchmark study rated medium for reproducibility. SMELT (middle author) reports compute-matched scaling laws for MoE looped transformers, claiming 6.8–18% compute savings under matched FLOPs, parameters, and KV-cache; the budget-matched evidence rests on single runs with proprietary models and is rated low for reproducibility. RePro, on proof-verified benchmark rewriting for reliable evaluation of LLM mathematical problem solving, is indexed but not yet reviewed. Topic signals (topic modeling, NLP techniques, neural network applications, AI planning) fall within the same computational scope. With 29 distinct coauthors against 3 recent works, collaboration scores 68 and activity 40 — a broad, recent, front-loaded record typical of an early-stage profile.
Collaboration pattern
The OpenAlex record lists 29 distinct coauthors across only 3 recent works — a wide per-paper network. Twelve collaborators recur, including Shaowen Wang, Ge Zhang, Jiaheng Liu, Wenhao Huang, and Shen Yan. Yuhao Wu takes first authorship on HarnessDev and middle positions on SMELT and RePro, suggesting both lead projects and contributions within larger teams.
Research keywords
Highlights
HarnessDev: LLMs Can Build Agent Harnesses, but Still Trail Human Systems
Compute-Matched Scaling Laws Claim 6.8–18% Savings for MoE Looped Transformers
RePro: Proof-Verified Benchmark Rewriting for LLM Math Evaluation
Co-authored 2026 arXiv work on rewriting LLM mathematical problem-solving benchmarks with proof verification for more reliable evaluation. It is indexed in the Evolve catalog but not yet reviewed — no verdict, keywords, or reproducibility rating exist yet.
Timeline
- 2026Publishes HarnessDev on arXiv as first author — a benchmark asking whether LLMs can create and evolve their own agent harnesses.
- 2026Co-authors SMELT, reporting compute-matched scaling laws for MoE looped transformers with claimed 6.8–18% compute savings.
- 2026Co-authors RePro on proof-verified benchmark rewriting for reliable evaluation of LLM mathematical problem solving (indexed, not yet reviewed).
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometrics are name-resolved from OpenAlex only; h- and i10-index read zero and no citation data is available, so output and impact are unscored.
- All catalogued works are 2026 arXiv preprints; no peer-reviewed venue or ORCID linkage is present.
- RePro is indexed but not yet reviewed by Evolve — no verdict, keywords, or reproducibility rating exist for it.
- Yuhao Wu is a common name; the profile may aggregate homonymous authors resolved through public bibliographic data.