Xue Liu

Mohamed bin Zayed University of Artificial Intelligence

Unclaimed · AI-structured profileOpenAlex

LLM evaluation researcher at MBZUAI building diagnostic benchmarks for time-series reasoning and role-play agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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-volume metrics (total papers, venues, output over time) exist in this snapshot. This axis would measure publishing volume and productivity; the two reviewed papers alone cannot support an estimate.

Impact—

Not yet scored: no citation counts, h-index, or uptake signals are available. This axis would measure how widely the work is cited and used; with only two 2026 arXiv preprints on file, influence cannot be estimated.

Focus0/100

The focus score is 0 on Evolve's 0-100 scale, and the raw field distribution explains why: the two reviewed papers split exactly 0.5 and 0.5 between 'LLM agents and role-play evaluation' and 'LLM evaluation for time-series reasoning'. Rather than one narrow line, the 2026 output forms a coherent evaluation agenda — diagnostic benchmarks probing whether LLMs truly understand — spread across two adjacent problem areas, which the metric reads as low concentration.

Collaboration55/100

The collaboration score is 55, built on 15 distinct coauthors across the reviewed record. Both 2026 papers are team efforts — TimeLitmus with a middle-author placement and ANIMASK with the final slot, neither corresponding — suggesting contribution within larger groups rather than solo or leading roles. Recurring partner names include Xiucheng Zhang, Zhuoning Xu, Hanjun Luo, Yankai Chen, and Hanan Salam, indicating repeat collaborators across projects.

Activity—

Not yet scored: no per-year publication cadence is available, so recent productivity trends are unknown. This axis would measure publishing pace and continuity over time; both catalog papers fall in 2026, which is too narrow a window to score.

Reproducibility55/100

Reproducibility averages 55 across the two papers assessed (reproPapers = 2), and Evolve rates each as 'medium' readiness. The limitation patterns are consistent and honest: TimeLitmus relies on rule-based labels over a small set of pairs, while ANIMASK reports single-run cells and author-side annotation. The shared profile — careful counterfactual designs undermined by small scale and limited runs — places both papers mid-range rather than at either extreme.

Research summary

Xue Liu is a researcher at the Mohamed bin Zayed University of Artificial Intelligence whose reviewed work centers on evaluating what large language models actually do, rather than on training new ones. Evolve's catalog holds two 2026 arXiv papers spanning two adjacent evaluation areas. TimeLitmus (middle author) introduces a diagnostic benchmark for event-conditioned time-series prediction, showing that LLMs' reported accuracy overstates true cross-modal understanding, checked through explanation faithfulness and counterfactual probes; Evolve flags rule-based labels and a small pair count as limitations. ANIMASK (final author) attributes role-play behavior in simulated story worlds to persona prompts versus model defaults using a freeze-replay design; Evolve finds the attribution convincing and honest while noting single-run cells and author-side annotation. This balanced split — 0.5 versus 0.5 across the two fields — yields a focus score of 0 on Evolve's 0-100 scale, and a collaboration score of 55 reflects 15 distinct coauthors. Reproducibility averages 55 across the two assessed papers, both rated medium readiness. Output, impact, and activity metrics are not yet available, so productivity and citation influence cannot be characterized. Because 'Xue Liu' is a common name, this profile is name-resolved from public bibliographic data and should be read with that caveat.

Collaboration pattern

Collaboration is mid-scale (score 55) around 15 distinct coauthors, and both 2026 papers are group efforts with non-leading positions — middle author on TimeLitmus, final author on ANIMASK, neither corresponding. Recurring names include Xiucheng Zhang, Zhuoning Xu, Hanjun Luo, Yankai Chen, Hanan Salam, Jie Gong, Maowei Jiang, Zhiwei Liu, Guojun Xiong, and Min Peng, suggesting repeat partners; 'Xue Liu' also appears inside this co-author list, a likely same-name artifact worth verifying.

Research keywords

LLM evaluationdiagnostic benchmarkevent-conditioned time-series predictioncross-modal understandingexplanation faithfulnesscounterfactual evaluationshortcut learningpersona attributionrole-play evaluationLLM agents and simulationcounterfactual persona removalsimulated story worlds

Highlights

TimeLitmus: probing whether LLMs truly understand time series when events intervene
A 2026 arXiv diagnostic benchmark for event-conditioned time-series prediction. Evolve's verdict: solid — LLM accuracy on event-conditioned forecasting overstates genuine cross-modal skill, exposed through explanation-faithfulness and counterfactual checks; rule-based labels and a small pair count temper the claim. Xue Liu is a middle author.
ANIMASK: separating persona prompts from model defaults in role play
A 2026 arXiv study using a freeze-replay design to attribute role-play behavior in simulated story worlds to persona instructions versus model defaults. Evolve finds the attribution convincing and honest, with single-run cells and author annotation as caveats. Xue Liu is the final author, not corresponding.

Timeline

2026TimeLitmus, a diagnostic benchmark for cross-modal understanding and explanation faithfulness in event-conditioned time-series prediction, appears on arXiv with Xue Liu as a middle author.
2026ANIMASK, attributing role-play in simulated story worlds to persona versus model defaults via freeze-replay, appears on arXiv with Xue Liu in the final author position.
2026Both reviewed papers land in the same year across two evaluation areas, with 15 distinct coauthors recorded and an average reproducibility score of 55 across the two papers.

Papers in the Evolve catalog

Frequent co-authors

Xiucheng Zhang×1 Zhuoning Xu×1 Hanjun Luo×1 Yankai Chen×1 Hanan Salam×1 Xue Liu×1 Jie Gong×1 Maowei Jiang×1 Zhiwei Liu×1 Yankai Chen×1 Guojun Xiong×1 Min Peng×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.

Dimension scores and raw metrics come from OpenAlex-derived public bibliometric data only; no citation counts, publication totals, or per-year activity were available.
'Xue Liu' is a common name; this profile is name-resolved from public bibliographic data, and the same name also appears in this profile's own co-author list, suggesting a same-name or self-listing artifact.
Coverage is limited to two 2026 arXiv preprints reviewed by Evolve; earlier or venue-published work is not represented.
Neither catalog paper lists this researcher as corresponding author, so research leadership cannot be assessed.