Early-stage first and corresponding author of 2026 arXiv work on LLM safety benchmarking and multi-agent conformity
2
papers here
2
reviewed by Evolve
2
first-author
2
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
Insufficient data: this snapshot contains only two cataloged arXiv papers and no full publication count or venue record. This axis would measure total scholarly output over time — paper count and venue distribution — and is left unscored until a complete publication history is available.
Impact—
No citation counts or downstream-usage records are available in this snapshot. This axis would measure the citation-based influence and community uptake of the researcher's work; until citation data accumulate for the 2026 papers, it cannot be honestly scored.
Focus0/100
Scores 0 on the focus axis. The raw field distribution is an even split — 0.5 natural language processing (AI safety / trustworthy ML) and 0.5 machine learning (LLM multi-agent systems) across the two cataloged papers. Both works target LLM trustworthiness — benchmarking moderators versus filtering peer conformity — so the zero reflects a deliberately two-sided lens within one theme, yet the metric records no single dominant field.
Collaboration14/100
Scores 14 on the collaboration axis. The raw record shows 1 distinct co-author, and the frequent-coauthor list contains only Yibo Hu's own name, consistent with solo first-and-corresponding authorship on both 2026 papers. No external collaboration network, lab, or cross-institution tie is visible in this snapshot, so collaborative behavior beyond independent work cannot yet be assessed.
Activity—
Only two papers, both dated 2026, appear in the reviewed sample, so year-by-year productivity cannot be computed. This axis would measure publication cadence and sustained research activity over time; it stays unscored until more years of data exist.
Reproducibility100/100
Scores 100 — the axis maximum — computed as a mean of 100 across 2 assessed papers. Both cataloged works, Safety-Flag and One Axis, No Brake, carry high reproducibility readiness, and the benchmark paper is openly released. A perfect score signals strong artifact and protocol disclosure habits, though it rests on a small two-paper sample.
Research summary
Yibo Hu is an early-stage researcher whose public bibliographic footprint, name-resolved from bibliometric data, currently centers on two first-and-corresponding-authored arXiv papers from 2026, both on the safety and trustworthiness of large language models. The first, Safety-Flag, unifies 7 safety benchmarks into one balanced flag protocol for LLM content moderators and is openly released, showing that error direction, calibration, and abstention diverge across LLM moderators. The second, One Axis, No Brake, is a clean, well-supported negative result: filtering harmful peer revisions in LLM groups reduces to imperfect self-knowledge, with AUROC between 0.64 and 0.89. The field distribution splits evenly between natural language processing (AI safety / trustworthy ML) and machine learning (LLM multi-agent systems), a two-sided but thematically coherent agenda. The recorded collaboration footprint is minimal — 1 distinct co-author, with only the researcher's own name recurring — consistent with solo execution of benchmarking, experiments, and writing. Reproducibility is the standout strength: a mean score of 100 across 2 assessed papers, both rated high readiness. Output, impact, and activity metrics are not yet computable from this snapshot, so the profile should be read as a promising early signal rather than an established record. No institution, country, or ORCID is attached to this profile.
Collaboration pattern
The snapshot shows an essentially solo working pattern: 1 distinct co-author is recorded and the frequent-coauthor list contains only Yibo Hu's own name. Both 2026 papers are first-authored and corresponding-authored, suggesting independent handling of benchmark construction, experiments, and writing. No recurring external collaborators, labs, or institutional ties are visible yet, so collaboration behavior beyond solo work cannot be characterized from these data.
Research keywords
Highlights
Unifying LLM moderator evaluation into one balanced flag protocol
A clean negative result: self-knowledge cannot filter harmful peer conformity
Timeline
- 2026Publishes Safety-Flag on arXiv as first and corresponding author: 7 safety benchmarks unified into one balanced flag protocol for LLM content moderators, openly released and rated high reproducibility readiness.
- 2026Publishes One Axis, No Brake, again as first and corresponding author, showing that filtering harmful peer conformity in LLMs reduces to imperfect self-knowledge (AUROC 0.64–0.89); also rated high reproducibility readiness.
- 2026Both cataloged papers land in the same year, splitting evenly between NLP/AI safety and ML/multi-agent systems, with a reproducibility mean of 100 across 2 assessed papers — an early, artifact-forward research agenda.
Papers in the Evolve catalog
- first authorcorresponding
One Axis, No Brake: Self-Knowledge Limits the Filtering of Harmful Peer Conformity in LLMs
- first authorcorresponding
Safety-Flag: A Unified Benchmark for the Reliability and Calibration of LLM Content Moderators
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 scores are computed from OpenAlex-based public data; only the two cataloged arXiv (2026) papers were reviewed for this profile.
- No institution, country, or ORCID is provided; the profile is name-resolved from public bibliographic data and may be ambiguous for a common name.
- Output, impact, and activity dimensions lack underlying data (no citation counts, venue record, or multi-year publication history) and are left unscored.
- The frequent-coauthor list contains only the researcher's own name, so the collaboration axis rests on a single distinct co-author record.