Jiawei Zhang

New York University

Unclaimed · AI-structured profileOpenAlex

New York University researcher working across LLM hallucination interpretability and online optimization theory

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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: no publication-count or productivity data is available beyond the two reviewed catalog papers. This axis would measure total publication volume, venue spread, and per-year output counts, none of which are in the current snapshot.

Impact—

Not yet scored: citation counts, h-index, and citation-based influence are absent from this snapshot. The impact axis would measure how widely the two 2026 preprints and any other work are cited, which cannot be estimated yet.

Focus0/100

The focus score is 0 because the reviewed catalog splits evenly across two unrelated fields: 0.5 natural language processing (LLM hallucination detection and interpretability) and 0.5 online optimization / operations research. Neither 2026 paper concentrates the profile in a single area, so this reads as a deliberately two-sided portfolio rather than a specialized one — a score that may shift as more papers are added.

Collaboration22/100

The collaboration score is 22, derived from only 2 distinct coauthors in the snapshot — Xuhan Tong and Menglong Li — who span both reviewed papers. Both 2026 papers place Jiawei Zhang in the last author position with no corresponding-author flag, a pattern consistent with a small, stable team and a senior or supervisory seat rather than a broad collaboration network.

Activity—

Not yet scored: this snapshot lacks publication-cadence data such as submission dates or yearly output beyond the shared 2026 stamp on both papers. The activity axis would measure research velocity and recent productivity, which cannot be assessed from the current data.

Reproducibility55/100

The reproducibility score is 55, a mean computed over 2 rated papers. Both carry a medium readiness rating: the hallucination work reports a concrete 0.910 AUROC but is confined to 7-8B models with supervised training, while the allocation paper is judged theoretically sound yet ships without experiments and with potentially large constants.

Research summary

Jiawei Zhang is a researcher at New York University whose reviewed work spans two distinct areas: natural language processing — specifically LLM hallucination detection and interpretability — and online optimization / operations research. The Evolve catalog holds two 2026 arXiv papers, both with Zhang in the last author position and neither flagged as corresponding author. The first ties confident LLM hallucinations to entity-keyword binding and introduces a keyword-perturbation detector built on causal-intervention signals, reaching 0.910 AUROC, though its scope is limited to 7-8B models and supervised training. The second proposes a first-order learning algorithm for online resource allocation with constant regret, removing the need for LP solves or nondegeneracy assumptions; the theory is judged sound, but constants may be large and experiments are absent. Bibliometrically the snapshot is narrow: 2 distinct coauthors (Xuhan Tong, Menglong Li), a field distribution split evenly at 0.5/0.5 between the two areas (focus score 0), a collaboration score of 22, and a reproducibility mean of 55 across 2 papers, both rated medium readiness. Output, impact, and activity scores are missing because the underlying data is insufficient, making this an early-stage profile.

Collaboration pattern

Collaboration is narrow and stable: 2 distinct coauthors — Xuhan Tong and Menglong Li — span the two reviewed 2026 papers, yielding a collaboration score of 22. Jiawei Zhang sits in the last author position on both papers without a corresponding-author flag, consistent with a senior or supervisory role within a small team. No cross-institution or large-consortium patterns are visible in the available data.

Research keywords

hallucination detectionlarge language modelskeyword perturbationcausal interventionentity-keyword bindingpretraining frequencyonline resource allocationnetwork revenue managementconstant regretfirst-order methodsnondegeneracyonline linear programming

Highlights

Confident hallucinations traced to entity-keyword binding
The 2026 study links confident LLM hallucinations to how models bind entity keywords, using keyword perturbation and causal intervention to probe the mechanism, with pretraining frequency identified as a contributing factor.
Keyword-perturbation detector reaches 0.910 AUROC
The proposed detector achieves 0.910 AUROC for hallucination detection, but its scope is limited to 7-8B-parameter models and relies on supervised training, which bounds its out-of-the-box applicability.
First-order policy with constant regret for online allocation
A first-order learning algorithm attains O(1) regret for online resource allocation without LP solves or nondegeneracy assumptions; the theory is judged sound, though constants may be huge and experiments are absent.

Timeline

2026Published 'When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs' on arXiv, in the last author position.
2026Published 'A First-Order Learning Algorithm for Online Resource Allocation with Constant Regret' on arXiv, also as last author.
2026Both catalog papers land in the same year, splitting the profile evenly between LLM hallucination research and online optimization (0.5/0.5 field distribution).

Papers in the Evolve catalog

Frequent co-authors

Xuhan Tong×1 Menglong 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.

'Jiawei Zhang' is a common name; this profile is name-resolved from public bibliographic data plus the two Evolve-reviewed arXiv papers.
Metrics cover only the Evolve-reviewed snapshot (OpenAlex-scope bibliometrics); citation counts, h-index, venue list, country, and ORCID were unavailable.
Output, impact, and activity radar axes are missing because the underlying data is insufficient — no scores are estimated for them.
Both papers are 2026 arXiv preprints; peer-review status and final publication venues are not reflected.