University of Liverpool researcher probing cost-efficient LLM agent debate and multimodal evidence-order evaluation
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 data is available yet, so this axis is left unscored. It would measure the size and cadence of the record — total papers, venue types, and yearly output — none of which can be inferred from the two 2026 preprints here.
Impact—
Citation-based impact is not yet in the snapshot, so the axis stays unscored. It would aggregate citations, citing works, and any h-index-like measures; for brand-new 2026 preprints these signals would in any case take time to accumulate.
Focus0/100
Evolve records an even split across two fields — 'multimodal large language model evaluation' and 'artificial intelligence (LLM multi-agent systems)', each at 0.5 of the field distribution — which yields a focus score of 0 on the 0-100 scale. With only two catalogued papers, one per field, the split reflects a dual-track early agenda rather than a settled specialisation; more publications will show which line consolidates.
Collaboration32/100
The collaboration score is 32 out of 100, based on 4 distinct co-authors in the current snapshot. The recurring names are Zhuoyun Li, Boxuan Wang, and Xiaowei Huang, and Yi Dong sits in the last author position on both catalogued papers without corresponding-author status — a pattern consistent with supervising a small, stable team of two or three co-workers rather than joining large consortium-style projects.
Activity—
Annual activity metrics are unavailable, so this axis is not scored. It would track output rhythm — papers per year, gaps, and recent momentum — data the current two-paper snapshot, both dated 2026, cannot support.
Reproducibility38/100
Reproducibility scores 38 out of 100, computed as a mean of 38 across the 2 papers with reproducibility assessments. Readiness is uneven: the multi-agent debate paper is rated medium, while the multimodal evidence-order paper is rated low, and the debate preprint is flagged as releasing no code. Consistent artefact release would be the clearest lever for lifting this axis.
Research summary
Yi Dong is a researcher at the University of Liverpool whose Evolve profile currently rests on two 2026 arXiv preprints spanning two adjacent areas of large language model research: LLM multi-agent systems and multimodal large language model evaluation, each accounting for 0.5 of the field distribution. On both papers Dong appears in the last author position without corresponding-author status, alongside a small set of four distinct co-authors — with Zhuoyun Li, Boxuan Wang, and Xiaowei Huang recurring — a pattern consistent with leading a compact, stable group. The multi-agent systems work asks whether complex topology control is needed at all in sparse multi-agent debate: it argues that random routing to distinct peers plus a round-three stopping rule rivals elaborate debate topologies at lower token cost, though Evolve notes the accuracy gains fall within seed noise and no code accompanies the paper. The evaluation work shows that presenting the same evidence in different orders shifts multimodal judgments toward image or speech, a carefully controlled empirical result whose 'noncommutativity' framing Evolve judges slightly broader than its two-element experiments. Radar scores reflect the early stage of this record: collaboration 32 and reproducibility 38 (a mean across two assessed papers), focus 0 given the even two-field split, while output, impact, and activity axes remain unscored for lack of data.
Collaboration pattern
The record shows a compact collaboration footprint: 4 distinct co-authors, giving a collaboration score of 32/100. Zhuoyun Li, Boxuan Wang, and Xiaowei Huang recur across the catalog, and Yi Dong takes the last author position on both papers without corresponding-author status — consistent with supervising or mentoring a small, stable group rather than joining large multi-institution consortia. All current output is arXiv preprints, so no venue-based co-author networks are yet visible.
Research keywords
Highlights
Sparse multi-agent debate without complex topology control
Evidence order sways multimodal judgment
Timeline
- 2026Publishes an arXiv preprint arguing that distinct-peer random routing and round-three stopping make complex topology control unnecessary in sparse multi-agent debate, in the last author position.
- 2026Publishes an arXiv evaluation preprint showing that reordering identical evidence shifts multimodal judgments toward image or speech, again as last author.
- 2026Both 2026 preprints enter the Evolve catalog, splitting the record evenly between LLM multi-agent systems and multimodal LLM evaluation (0.5 each in the 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.
- Radar dimension scores are computed from public bibliometric records only; no ORCID or country field is available to confirm identity, and 'Yi Dong' is a relatively common name — this profile is name-resolved from public bibliographic data.
- The output, impact, and activity axes are unscored because the underlying publication counts, citations, and yearly activity metrics are not in the current snapshot.
- The profile rests on only two papers, both 2026 arXiv preprints in the Evolve catalog; no peer-reviewed venues, grants, or awards are included.
- The reproducibility mean (38) covers only 2 assessed papers, one rated medium and one low; the multi-agent debate preprint is flagged as releasing no code.