Princeton University researcher in LLM evaluation, agent benchmarks, and compute-efficient model architectures
6
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
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—
Not scored: the underlying publication-volume data for this axis is insufficient in the current snapshot. This dimension would measure total indexed works, per-year output counts, and authorship positions; only partial counts (22 works in three years) are visible so far.
Impact—
Not scored: citation data is too sparse to characterize influence (h-index 1, i10-index 1). This dimension would measure citation counts, h-index trajectory, and downstream uptake; a meaningful score needs more citing literature to accumulate.
Focus100/100
Scores 100/100. OpenAlex topic assignments concentrate almost entirely on language-centered AI: topic modeling (17 records) and natural language processing techniques (14) dominate, with AI-based problem solving and planning (6) trailing and only minor counts in reinforcement learning in robotics (3) and multimodal machine learning (3). The catalog matches — LLM benchmarks, self-evolving agents, and transformer architecture — consistent with the single recorded field, artificial intelligence (LLM evaluation).
Collaboration88/100
Scores 88/100. The OpenAlex record lists 83 distinct co-authors, an unusually wide network for so early a profile. Most 2026 catalog papers place Huang in a middle author position, with one last-author slot (xDailyBench). Recurring collaborators include Jiajun Shi, Jiaheng Liu, Ge Zhang, Shen Yan, Xinping Lei, Zexuan Wang, and Jingyuan Zhang, indicating dense team-based preprint collaboration rather than solitary work.
Activity91/100
Scores 91/100. Indexed works accelerate sharply across the provided years: 1 in 2024, 4 in 2025, and 17 in 2026, with 22 works in the last three years — essentially the entire visible record compressed into a steep recent climb. All six Evolve-catalog papers are 2026 arXiv preprints, confirming an intensive current publishing cadence; the caveat is that preprint-heavy output is not yet filtered by peer review.
Reproducibility55/100
Scores 55/100, but the signal is thin: it is computed from only 1 reviewed paper (reproMean 55, reproPapers 1). Within the catalog, xDailyBench and HarnessDev rate medium reproducibility — the former praised for solid rubric design yet flagged for missing variance and bias checks — while SMELT rates low, being single-run, proprietary, and unverifiable. The spread suggests inconsistent artifact and evaluation transparency rather than an established openness practice.
Research summary
Wenhao Huang is a Princeton University researcher whose catalog centers on evaluating large language models and agentic systems, alongside model architecture research. The OpenAlex-derived snapshot shows an early-stage profile in a steep growth phase: 22 works in the last three years, rising from 1 indexed work in 2024 to 4 in 2025 and 17 in 2026, while citation indicators (h-index 1, i10-index 1) have not yet caught up. Topic assignments concentrate on topic modeling (17 records) and natural language processing techniques (14), with smaller footprints in AI-based problem solving and planning (6), reinforcement learning in robotics (3), and multimodal machine learning applications (3). Six 2026 arXiv papers appear in the Evolve catalog. xDailyBench, where Huang is last author, introduces a 248-task benchmark of real professional consultation requests, finding that models trail explicit requests by at least 9 points on implicit needs across 11 models. HarnessDev asks whether LLMs can build and evolve their own agent harnesses; SMELT studies compute-matched scaling laws for MoE looped transformers, reporting 6.8-18% compute savings. Aspire, S3Gym, and REER-PT extend the self-evolution and training-data agenda. With 83 distinct co-authors and recurring collaborators such as Jiajun Shi, Jiaheng Liu, Ge Zhang, and Shen Yan, the work pattern is team-based and prolific.
Collaboration pattern
The OpenAlex record lists 83 distinct co-authors across 22 works in three years, an unusually wide network for an early-stage profile. Most 2026 catalog papers place Huang in a middle author position, with one last-author slot (xDailyBench) and no corresponding-author entries. Recurring names include Jiajun Shi, Jiaheng Liu (appearing twice in the frequency data), Ge Zhang, Shen Yan, Xinping Lei, Zexuan Wang, Jinkai Liu, Yuhao Wu, Yuxuan Zhang, Yongchang Peng, and Jingyuan Zhang — consistent with team-based benchmark and systems projects.
Research keywords
Highlights
248-task benchmark of real professional consultation requests
Can LLMs build and evolve their own agent harnesses?
Scaling laws for compute-matched MoE looped transformers
A 2026 cluster on self-evolving and self-improving LLMs
Three further 2026 preprints — Aspire (self-evolution from vague goals), S3Gym (self-testing and self-judging), and REER-PT (perplexity-guided pre-training data augmentation) — extend the same self-improvement agenda, though Evolve has not yet reviewed them.
Timeline
- 2024First indexed work appears in the OpenAlex record, with 1 work counted for the year.
- 2025Indexed output grows to 4 works.
- 2026Output jumps to 17 indexed works; six arXiv preprints enter the Evolve catalog, spanning LLM benchmarks, agent harnesses, self-evolution, and MoE looped-transformer scaling laws.
- 2026xDailyBench, with Huang in the last author position, is reviewed by Evolve as a 248-task benchmark of real professional consultation.
Papers in the Evolve catalog
xDailyBench: Benchmarking LLMs on Professional Consultation for Real-Life Problems
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
Aspire: Can Models Self-Evolve from Vague Goals?
S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?
REER-PT: Reverse-Engineered Reasoning for Perplexity-Guided Pre-training Data Augmentation
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.
- Bibliometric indicators (h-index, i10-index, counts by year, topics, co-author counts) come from OpenAlex only; no ORCID is linked, so identity is name-resolved from public bibliographic data.
- Output and impact axes are unscored because the underlying citation and coverage data is insufficient; h-index and i10-index stand at 1.
- Reproducibility scoring covers only 1 reviewed paper; Aspire, S3Gym, and REER-PT lack Evolve verdicts, keywords, and reproducibility ratings.
- All reviewed papers are 2026 arXiv preprints; peer-review status and venue acceptance are not reflected in this data.