UC Berkeley researcher building rigorous anti-gaming benchmarks for LLM agent deployment and post-training delivery.
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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 yet scored: this snapshot lacks the publication-volume data the output axis needs, so no estimate is given. The axis would measure total paper production and how consistently outputs appear across years.
Impact—
Not yet scored: citation counts and influence indicators are unavailable in this snapshot, so the impact axis is left blank. It would measure how widely the work is cited and adopted within LLM-agent evaluation.
Focus0/100
The field distribution is exactly 50% software engineering (LLM agent evaluation) and 50% machine learning (LLM agent evaluation), yielding a focus score of 0 on the platform's concentration scale. In substance the portfolio is tightly themed: both reviewed papers target LLM-agent evaluation, one on deployment-environment configuration and one on post-training delivery, so the low score reflects an even split between two field labels rather than scattered topics.
Collaboration36/100
The bibliometric record shows 5 distinct coauthors and a collaboration score of 36. Both reviewed papers are first-authored and non-corresponding, with Junfei Zhan, Yueting Li, and Qirong Guo recurring among frequent coauthors, pointing to a small, stable core team around a first-author-driven benchmark agenda rather than a large distributed network.
Activity—
Not yet scored: recency-weighted publication cadence is not available in this snapshot, so activity cannot be estimated. The axis would measure how recently and how steadily the researcher continues to publish.
Reproducibility78/100
Reproducibility is the strongest scored axis: the mean score across 2 assessed papers is 78. FDE-Bench carries high repro-readiness, consistent with its replay-graded design and anti-gaming gates, while the post-training delivery benchmark is rated medium. Together these point to a habit of building released, checkable evaluation infrastructure.
Research summary
Weihang Ding is a researcher at the University of California, Berkeley whose work sits at the intersection of software engineering and machine learning, centered on evaluating large language model (LLM) agents in realistic operational roles. The Evolve catalog lists two 2026 arXiv papers, both first-authored: FDE-Bench, a replay-graded benchmark for LLM agents that configure deployment environments, with Docker and Kubernetes among the tested settings; and 'Trains but Doesn't Learn,' a post-training delivery benchmark that frames LLM agents as forward-deployed engineers, reporting 0 silent failures across 38 runs while 31 of 38 runs fail contract gates. Evolve's verdicts call both papers rigorous, crediting FDE-Bench with strong anti-gaming gates, while flagging single-episode cells and an n=1 human baseline in the first and a cross-grid human comparison in the second. Platform bibliometrics show a 50/50 split between software engineering (LLM agent evaluation) and machine learning (LLM agent evaluation), 5 distinct coauthors, and a mean reproducibility score of 78 across 2 assessed papers, with high and medium repro-readiness respectively; frequent coauthors include Junfei Zhan, Yueting Li, and Qirong Guo. Output, impact, and activity axes are not yet scored because the underlying data is insufficient.
Collaboration pattern
The visible pattern is a compact, first-author-led collaboration: 5 distinct coauthors and a collaboration score of 36, with first (non-corresponding) authorship on both reviewed papers. Junfei Zhan, Yueting Li, and Qirong Guo recur as frequent coauthors, suggesting a small stable team that ships benchmark artifacts together. The source coauthor list also contains a duplicate and the researcher's own name, a resolution artifact; no senior-lab or cross-institution ties are visible in this snapshot.
Research keywords
Highlights
FDE-Bench: a replay-graded, anti-gaming benchmark for LLM deployment agents
'Trains but Doesn't Learn': governing post-training delivery by LLM agents
Timeline
- 2026Publishes FDE-Bench on arXiv as first author: a replay-graded benchmark for LLM agents configuring deployment environments, with Docker and Kubernetes in scope.
- 2026Publishes 'Trains but Doesn't Learn' on arXiv as first author, benchmarking LLM agents as forward-deployed engineers on governed post-training delivery (0 silent failures in 38 runs; 31/38 fail contract gates).
- 2026Both papers enter the Evolve catalog with reproducibility assessments — high readiness for FDE-Bench, medium for the delivery benchmark — yielding a mean reproducibility score of 78 across 2 papers.
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.
- Bibliometric signals derive from OpenAlex-based public bibliographic data; the output, impact, and activity axes lack sufficient underlying data and are not scored.
- Profile is name-resolved from public bibliographic data and no ORCID is linked, so identity confidence is limited.
- Coverage is limited to the two Evolve-reviewed 2026 arXiv papers; any other publications fall outside this snapshot.
- The frequent-coauthor list contains a duplicate entry and the researcher's own name, indicating entity-resolution noise in the source data.