Microsoft Research (UK) researcher working across LLM red teaming and LLM multi-agent systems.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
Output metrics — publication volume, productivity trends, and venue mix over time — are not yet computable for this profile because the underlying bibliographic coverage is insufficient; the axis is therefore left unscored rather than estimated.
Impact—
Impact measures — citation counts, citation trajectory, and downstream uptake of the work — are not available in the current data snapshot; this axis would capture the influence of published research, and is left unscored pending sufficient coverage.
Focus0/100
Focus scores 0, the lowest measured concentration: the reviewed output splits evenly, 0.5 to AI safety / LLM red teaming and 0.5 to LLM multi-agent systems. With only two reviewed papers, both from 2026, this even split reads as two parallel LLM directions rather than a single specialized line; the score reflects measured topical spread in a very small sample, not an absence of domain expertise.
Collaboration59/100
Collaboration scores 59, computed from 18 distinct coauthors in the available bibliographic record, with 12 recurring collaborators including Li Dong, Furu Wei, Dongdong Zhang, and Tengchao Lv. On both cataloged 2026 papers the researcher sits in a middle author position without corresponding-author role, consistent with a contributor-style pattern inside larger teams rather than small-group or solo-led work. The moderate score reflects a real but not exceptionally wide network in the sampled data.
Activity—
Activity measures — recent publication cadence and research tempo in recent years — cannot be computed from the data currently available; the axis tracks ongoing research output over time, and remains unscored until the underlying records are complete.
Reproducibility55/100
Reproducibility scores 55, the mean readiness rating across 2 reviewed papers, each rated medium. Evolve's verdicts identify specific gaps: CART's LLM-judged gains over static replay appear real but lack statistical rigor, while Agensh reports a +49% relative gain at 128 agents on ProgramBench with no variance, cost, or orchestrator-baseline reporting. The methods are described at a level that suggests implementability, but reported experimental detail does not yet support confident independent replication.
Research summary
Shaohan Huang is a researcher at Microsoft Research (United Kingdom) whose cataloged 2026 output spans two related corners of large language model research: AI safety through automated red teaming, and multi-agent systems through large-scale agent organization. The Evolve catalog holds two papers, both 2026 arXiv preprints with middle authorship and no corresponding role. CART introduces closed-loop adaptive red teaming for LLMs, using Thompson sampling and provenance tracking to move beyond static attack replay; Evolve judges the LLM-judged gains over static replay as looking real but lacking statistical rigor. Agensh reports an orchestrator-free harness scaling organizational intelligence to 1,024 agents, with a +49% relative gain at 128 agents on ProgramBench, though without variance, cost, or orchestrator-baseline reporting. Both papers carry a medium reproducibility rating, averaging 55. The bibliometric snapshot records 18 distinct coauthors with 12 recurring collaborators and a collaboration score of 59; the field distribution splits evenly, 0.5 to AI safety / LLM red teaming and 0.5 to LLM multi-agent systems, producing a focus score of 0. Output, impact, and activity metrics are not yet available and are left unscored. As Shaohan Huang is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Shaohan Huang's sampled record shows 18 distinct coauthors and 12 recurring collaborators, including Li Dong, Furu Wei, Dongdong Zhang, Tengchao Lv, Nan Yang, Lei Cui, Yupan Huang, Wenshan Wu, Xiangyang Zhou, Yilin Jia, Yuzhong Zhao, and Zhihao Zhan. On both cataloged 2026 papers the researcher contributes from a middle author position without corresponding-author role, suggesting work embedded in larger Microsoft Research teams rather than independently led projects; the collaboration dimension scores 59.
Research keywords
Highlights
Closed-loop adaptive red teaming for LLMs
Scaling organizational intelligence to 1,024 agents
Medium reproducibility across both 2026 papers
Both reviewed papers are rated medium on reproducibility readiness, giving a mean score of 55: strong method narratives paired with incomplete statistical and baseline reporting. This pairing of ambitious systems-level claims with medium verification readiness is the most consistent pattern in the current record.
A two-track 2026 output profile
The cataloged record splits evenly, 0.5 to AI safety / LLM red teaming (CART) and 0.5 to LLM multi-agent systems (Agensh), placing the researcher at the intersection of LLM safety evaluation and agentic systems rather than in a single niche.
Timeline
- 2026CART appears on arXiv, introducing closed-loop adaptive red teaming with Thompson sampling and provenance tracking for LLM safety evaluation.
- 2026Agensh appears on arXiv, reporting an orchestrator-free harness that scales to 1,024 LLM agents with a +49% relative gain at 128 agents on ProgramBench.
- 2026The two papers together define a two-track record spanning AI safety / LLM red teaming and LLM multi-agent systems, each at 0.5 of the cataloged field distribution.
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.
- Dimension scores are computed from a limited public bibliographic snapshot; the output, impact, and activity axes lack sufficient underlying data and are left unscored rather than estimated.
- The profile rests on only 2 reviewed papers in the Evolve catalog, both 2026 arXiv preprints; citation counts, venue acceptance, and career history are outside the current data scope.
- Shaohan Huang is a common name and no ORCID is provided; the profile is name-resolved from public bibliographic data, so identity confidence is limited.
- The reproducibility score of 55 is a mean over just 2 papers and should be read as an early-stage signal, not a stable estimate.