Researcher at USTC spanning LLM pretraining-data detection and spoken-dialogue role-play robustness benchmarking
2
papers here
2
reviewed by Evolve
0
first-author
0
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—
No publication-volume data is attached to this profile yet: the snapshot lacks paper counts and venue totals, so this axis is unscored. Once available, it would measure cumulative output, annual paper counts, and the spread of venues across the record.
Impact—
No citation data is present in the current snapshot — no citation counts, h-index, or per-paper impact figures — so this axis is unscored. It would measure how often this researcher's work is cited and built upon by later studies.
Focus0/100
The focus score is 0, and the underlying field distribution is an even 50/50 split between spoken dialogue systems (speech processing / NLP) and machine learning for LLM privacy and membership inference. The two reviewed papers sit one per field, so the record currently reads as two parallel lines rather than a single converged research agenda.
Collaboration51/100
Collaboration scores 51, built on 12 distinct coauthors in the snapshot. Both reviewed papers are multi-authored arXiv preprints: a middle-author position on the free-energy pretraining-data detection work and a last-author position on RoleBreak, neither as corresponding author — suggesting contributing rather than leading roles so far. Frequent coauthors on record include Yuqi Wang, Zhenya Huang, and Shijin Wang, among others.
Activity—
The snapshot carries no year-by-year publication cadence, so recent-productivity trends cannot be scored. This axis would measure annual output volume and whether research activity is accelerating, holding steady, or tapering in recent years.
Reproducibility100/100
Reproducibility scores 100, the maximum, with a reproMean of 100 across 2 papers — both reviewed works carry a high reproducibility-readiness rating. The free-energy detection paper reports small but consistent gains across 3 benchmarks, and RoleBreak contributes a 310-role benchmark, so both are built to be re-run; reviewers note only minor caveats (vacuous bounds; a single LLM judge).
Research summary
Qi Liu is a researcher at the University of Science and Technology of China whose reviewed work to date spans two parallel lines: machine learning for LLM privacy — pretraining-data detection and membership inference — and speech processing / natural language processing for spoken dialogue systems. Because 'Qi Liu' is an extremely common name, this is an early-stage profile name-resolved from public bibliographic data and should be read with that ambiguity in mind. The Evolve catalog holds two 2026 arXiv preprints. The first, 'Detecting Pretraining Data in Large Language Models from a Free-Energy Perspective,' scores membership via loss minus a λ-weighted entropy term; Evolve's review notes small but consistent gains across 3 benchmarks and neat theory with likely vacuous bounds. The second, 'RoleBreak,' benchmarks long-horizon role-playing robustness in spoken dialogue across 310 roles and finds spoken role-play breaking down by around turn 10, though a single LLM judge and absent error bars temper the headline claims. The field distribution is an even 50/50 split, consistent with a focus score of 0, while reproducibility scores 100 (reproMean over 2 papers, both rated high readiness). Collaboration scores 51 on 12 distinct coauthors, with both papers in non-corresponding positions. Output, impact, and activity metrics are not yet available in the underlying snapshot.
Collaboration pattern
The snapshot lists 12 distinct coauthors; both reviewed papers are multi-author arXiv preprints with Qi Liu in non-corresponding roles (middle author on the free-energy detection paper, last author on RoleBreak). Frequent coauthors on record include Yuqi Wang, Fengyuan Liu, Haochen Luo, Zhiqi Yu, Chenye Ke, Zirui Liu, Yan Zhuang, Jintao Zhang, Zhenya Huang, and Shijin Wang. The list contains duplicate 'Qi Liu' entries, likely self-matches from name resolution.
Research keywords
Highlights
Free-energy lens on pretraining-data detection
RoleBreak: spoken role-play breaks by around turn 10
Timeline
- 2026Free-energy-based pretraining-data detection for large language models appears on arXiv, with Qi Liu in a middle-author position.
- 2026RoleBreak, a benchmark for long-horizon role-playing robustness in spoken dialogue, appears on arXiv, with Qi Liu in the last-author position.
- 2026Both 2026 preprints enter the Evolve catalog rated high for reproducibility readiness, with a reproMean of 100 across the 2 papers.
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.
- All bibliometric scores and counts come from Evolve's bibliometric snapshot; no external sources were independently checked.
- 'Qi Liu' is an extremely common name; this profile is name-resolved from public bibliographic data and may mix records of different researchers.
- Only 2 papers are reviewed so far; output, impact, and activity axes lack underlying data (no publication counts or citation metrics).
- Both reviewed works are 2026 arXiv preprints; peer-review status and final venues are not captured.