Microsoft Research (UK) researcher across LLM red-teaming safety and large-scale 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—
No publication-volume or venue data is available in this snapshot, so an output score cannot be computed. This axis would measure the number, cadence and venue quality of the researcher's publications over time.
Impact—
Citation-based indicators are not available for this profile yet, so impact is left unscored. This axis would measure citations, citation velocity and scholarly uptake of the researcher's work relative to field norms.
Focus0/100
Focus scores 0 because the field distribution splits exactly evenly: 0.5 artificial intelligence (AI safety / LLM red teaming) and 0.5 LLM multi-agent systems (agentic AI). Rather than a single deep specialty, the portfolio divides evenly between two adjacent directions — adversarial safety testing of LLMs and decentralized multi-agent orchestration. The 50/50 split reads as an intentional two-track agenda rather than scattered breadth.
Collaboration59/100
Collaboration scores 59 on the strength of 18 distinct coauthors in the underlying bibliometric record, with 12 recurring collaborators — including Shaohan Huang, Furu Wei, Dongdong Zhang and Nan Yang — suggesting a stable core team. Both Evolve-reviewed papers place this researcher in a middle author position without corresponding-author role, consistent with a contributing member of larger group efforts rather than a solo or lead-corresponding profile.
Activity—
Temporal publishing-frequency data is insufficient in this snapshot, so activity cannot be scored. This axis would measure recent publication cadence — how actively the researcher is producing new work year over year.
Reproducibility55/100
Reproducibility averages 55 across the 2 papers with reproducibility assessments, both rated medium readiness in Evolve reviews. CART's Thompson-sampled closed-loop probing is described concretely enough to attempt, but its LLM-judged gains lack statistical rigor; Agensh reports a +49% relative gain at 128 agents on ProgramBench yet omits variance, cost and orchestrator baselines. Method descriptions are followable; evaluation evidence is the weak link.
Research summary
Li Dong is a researcher at Microsoft Research (United Kingdom) whose cataloged work spans two adjacent directions in modern AI: adversarial safety evaluation of large language models and large-scale LLM multi-agent systems. The Evolve catalog holds two 2026 arXiv papers, both with Li Dong as a middle author. CART introduces closed-loop adaptive red teaming built on Thompson-sampled, provenance-tracked probing; Evolve's review judges the LLM-judged gains over static replay as plausibly real but statistically under-supported. Agensh reports an orchestrator-free harness scaled to 1,024 LLM agents with a +49% relative gain at 128 agents on ProgramBench, though the review notes missing variance, cost and orchestrator baselines. Bibliometric indicators show 18 distinct coauthors, a collaboration score of 59 and roughly twelve recurring collaborators, pointing to a stable team context; the field distribution splits exactly 50/50 between AI safety / LLM red teaming and agentic AI, yielding a focus score of 0; reproducibility averages 55 across the two assessed papers, both rated medium readiness. Publication-volume, citation-impact and publishing-cadence data are absent from this snapshot, so those axes remain unscored rather than estimated. Because 'Li Dong' is a common name, this profile is name-resolved from public bibliographic data and should be read as an evolving record of the cataloged work rather than a verified biography.
Collaboration pattern
The co-author graph shows 18 distinct collaborators, with a recurring core of 12 — Shaohan Huang, Furu Wei, Dongdong Zhang, Tengchao Lv, Nan Yang and Lei Cui among them — indicating a stable lab-team pattern. On both Evolve-reviewed 2026 papers Li Dong sits in a middle author position and is not the corresponding author, consistent with contributing specialist work inside larger group projects rather than leading them. A collaboration score of 59 reflects moderate breadth with a strong stable core.
Research keywords
Highlights
Closed-loop, Thompson-sampled red teaming for LLMs
Orchestrator-free scaling to 1,024 LLM agents
Medium reproducibility across both reviewed papers
Both cataloged papers carry medium reproducibility readiness and average 55 on Evolve's reproducibility axis: methods are described concretely enough to attempt, but statistical controls and baseline comparisons are the weak spots.
Timeline
- 2026CART: Closed-Loop Adaptive Red Teaming for Large Language Models released on arXiv, introducing Thompson-sampled, provenance-tracked closed-loop probing.
- 2026Agensh released on arXiv, reporting an orchestrator-free harness scaled to 1,024 LLM agents with a +49% relative gain at 128 agents on ProgramBench.
- 2026Both papers entered the Evolve catalog with medium reproducibility readiness; the two-paper reproducibility mean is 55.
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.
- Radar scores and co-author counts derive from public bibliometric data (OpenAlex-sourced); no ORCID was available to verify identity.
- 'Li Dong' is a common name; this profile is name-resolved from public bibliographic records and may mix same-name authors.
- Output, impact and activity axes are unscored because publication-volume, citation and cadence data are absent from this snapshot.
- Highlights and the reproducibility mean rest on only two Evolve-reviewed 2026 arXiv preprints; peer-reviewed venue status is unverified.