Fei Wang

University of Science and Technology of China · China

Unclaimed · AI-structured profileOpenAlex

USTC researcher on LLM-agent reliability, world-model verification, and traceable robot-policy evaluation

papers here

reviewed by Evolve

first-author

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 reliable publication-count data for the profile — OpenAlex lists an h-index and i10-index of 0 — so this axis, which measures yearly paper output and productivity trend over time, stays unscored until de-duplicated records accrue.

Impact—

No score yet: citation counts, h-index and downstream uptake are missing from the current OpenAlex snapshot, so scholarly influence cannot be estimated; this axis would measure citations, referencing works and field-level uptake once such records accumulate.

Focus0/100

Scores 0 because effort is evenly split: the OpenAlex field distribution assigns exactly one-third each to reinforcement learning and LLM agents, robot learning / embodied AI evaluation, and LLM-agent AI, while the topic profile adds generative adversarial networks and image synthesis (4), multimodal applications (3), image enhancement (2), explainable AI (2) and topic modeling (1) — breadth across five-plus clusters rather than one concentrated line.

Collaboration64/100

Scores 64 on a broad but mid-weight pattern: the OpenAlex profile links 24 distinct coauthors in the recent window, with a recurring circle including Huatai Zhu, Qiang Chen, Ziqian Kou and Wenhao Li among roughly a dozen frequent names. All three reviewed 2026 papers place this researcher in the middle author position, never corresponding author — steady team participation rather than lone or lead-authored work.

Activity64/100

Scores 64 on a recent but compressed record: OpenAlex attributes 8 works to the last three years, and the counts-by-year series contains only 2026 (8 works), so the entire visible output lands in a single year. The three reviewed arXiv papers all carry 2026 timestamps, confirming an active current phase whose multi-year cadence cannot yet be judged.

Reproducibility70/100

Scores 70 as the mean repro-readiness across the 3 papers Evolve has reviewed: IndustrialVLA-Bench is rated high — a traceable, evidence-tiered six-model benchmark — while Dual-Frontier and ReAdapt are rated medium, the former for unclear risk calibration and the latter for a single-model result without ablations or significance tests. Overall, evaluation-heavy work with above-average replication transparency.

Research summary

Fei Wang is an early-career researcher at the University of Science and Technology of China whose recent work makes LLM-driven agents and robot policies measurable and trustworthy. Evolve's catalog holds three 2026 arXiv papers spread evenly across this profile's three fields: machine learning (reinforcement learning and LLM agents), robot learning / embodied AI evaluation, and artificial intelligence (LLM agents). 'Dual-Frontier' shows passive failure attribution is non-identifiable and motivates certified verify-then-promote gates for world models. 'IndustrialVLA-Bench' benchmarked six VLA/WAM models, finding clean LIBERO conditions near-saturated (1.6 pp spread) while robustness and paraphrase axes separate models by 15–31 pp. 'ReAdapt' reports explicit social-state updates lifting LLM-agent relational accuracy from 37% to 51% and 69% to 77%. OpenAlex attributes 8 works to the last three years — all landing in 2026 — with 24 distinct coauthors, and a topic profile weighted toward generative adversarial networks and image synthesis (4), multimodal machine learning applications (3), image enhancement (2) and explainable AI (2). That topic mix diverges from the agent-focused reviewed papers, a divergence plausibly reflecting same-name record fragmentation. Evolve's mean repro-readiness across the three papers is 70, with the benchmark rated high.

Collaboration pattern

Collaboration is broad and team-based: 24 distinct coauthors appear in the recent OpenAlex window, with a recurring circle including Huatai Zhu, Qiang Chen, Ziqian Kou, Wenhao Li and Yichao Cao. On all three reviewed 2026 papers Fei Wang sits in the middle author position and is never the corresponding author, suggesting a contributor role within larger groups. Notably, the frequent-coauthor list itself contains a second 'Fei Wang', hinting at name-collision noise in the underlying graph.

Research keywords

LLM agentsworld modelsfailure attributiondecision reliabilityconformal predictionvision-language-action modelsrobot manipulation benchmarkingrobustness evaluationsocial reasoningexplainable AIgenerative adversarial networksmultimodal machine learning

Highlights

Certified verify-then-promote gates for agent world models
Evolve's review calls it strong theory: passive failure attribution is shown to be non-identifiable, motivating certified verify-then-promote gates; benchmarks are strong but risk calibration remains unclear (repro readiness: medium).
Six-model VLA benchmark shows clean-task saturation
The traceable multi-axis benchmark finds clean LIBERO conditions near-saturated at a 1.6 pp spread while robustness and instruction-paraphrase axes separate six VLA/WAM models by 15–31 pp; Evolve rates repro readiness high, though on n=3 runs.
Structured social-state updates lift LLM-agent relational reasoning
ReAdapt reports that explicit social-state updates raise LLM-agent relational accuracy from 37% to 51% and 69% to 77%; Evolve views it as a credible proof-of-concept limited by a single model and no ablation or statistical tests (repro readiness: medium).

Timeline

2026Publishes 'Dual-Frontier: When Can an Agent Trust Its World Model?' on arXiv, establishing the non-identifiability of passive failure attribution and proposing certified verify-then-promote gates.
2026Releases IndustrialVLA-Bench, a traceable multi-axis evaluation of six open robot policy models across clean LIBERO, robustness and instruction-paraphrase axes.
2026Contributes to ReAdapt, showing structured social-state updates lifting LLM-agent relational accuracy from 37% to 51% and 69% to 77%.
2026OpenAlex attributes 8 works to 2026 — the only year in the counts-by-year series — with 24 distinct coauthors across the recent window.

Papers in the Evolve catalog

Frequent co-authors

Fei Wang×2 Huatai Zhu×1 Qiang Chen×1 Ziqian Kou×1 Wenhao Li×1 Yichao Cao×1 Xiu Su×1 Yi Chen×1 Yiqi Wang×1 Zhifeng Rao×1 Jiaqi Zhang×1 Xiaoyang Li×1

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 come from OpenAlex only; the snapshot reports an h-index and i10-index of 0, and the output and impact axes are marked insufficient, so citation-based standing is unassessed.
'Fei Wang' is a very common name and no ORCID is provided; this profile is name-resolved from public bibliographic data, and the frequent-coauthor list even contains a second 'Fei Wang', indicating possible identity fragmentation.
OpenAlex topics (GANs and image synthesis, image enhancement, XAI, topic modeling) diverge from the reviewed 2026 papers on LLM agents and robot evaluation, so the topic mix may blend records from different same-name authors.
All three reviewed papers are 2026 arXiv preprints; peer-review status and citation uptake are unknown.