Qi Liu

University of Hong Kong

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher at the University of Hong Kong working on LLM coding agents and pretraining-data privacy.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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—

The underlying data for this axis (publication volume and author-role distribution over time) is not yet available; it would measure the size and cadence of Liu's output, including authorship positions across venues, and cannot be estimated from the current record.

Impact—

No citation or influence metrics are available yet; this axis would measure how often Liu's papers are cited and adopted by follow-up work, which cannot be estimated from two recent 2026 preprints in the reviewed catalog.

Focus0/100

The focus score is 0, reflecting an even 50/50 split between two subfields — LLM coding agents and LLM privacy/membership inference — rather than one concentrated line. The two reviewed 2026 papers sit in these areas respectively, suggesting breadth across LLM themes but no dominant specialty yet; the score may change as more publications accumulate in one area.

Collaboration68/100

Liu scores 68 on collaboration, based on 29 distinct coauthors in the reviewed record and 12 recurring collaborators, including Bowen Ye, Lei Li, Shicheng Li, and Zihao Yue. Both 2026 papers were written in middle authorship positions within sizeable teams, indicating a cooperative working style rather than solo research, though neither paper lists Liu as corresponding author.

Activity—

Temporal activity data (publications per year) is insufficient right now; this axis would measure research cadence and continuity over time, which cannot be inferred from the current data since both reviewed papers share the single publication year 2026.

Reproducibility60/100

Reproducibility scores 60, the mean across 2 scored papers whose readiness differs sharply: the free-energy pretraining-data detection paper is rated high, while CodeMidas is rated low because its reported gains are single-run without significance tests. The mixed profile suggests solid methodological reporting in one line and claims still awaiting verification in the other.

Research summary

Qi Liu is a machine learning researcher at the University of Hong Kong whose reviewed work concentrates on two complementary fronts of large language model research: agentic coding with reinforcement learning, and privacy-oriented pretraining-data detection. The evidence base is small — two 2026 arXiv preprints reviewed by Evolve, on neither of which Liu is first or corresponding author — and the profile is name-resolved from public bibliographic data because 'Qi Liu' is a common name. In CodeMidas, the team claims that source code alone can be converted into 5,545 reinforcement learning environments that lift MiMo-V2.5 across five benchmarks; Evolve judged the gains plausible but flagged single-run results without significance tests and low reproduction readiness. A companion line proposes an ETD score — model loss minus a lambda-weighted entropy term, framed via Helmholtz free energy — for detecting pretraining data, with small but consistent gains across three benchmarks and high reproduction readiness. Bibliometric signals show a collaborative, split-focus profile: 29 distinct coauthors, 12 recurring collaborators, a collaboration score of 68, an even 50/50 field split, and a mean reproducibility score of 60 across 2 scored papers. Output, impact, and activity metrics are not yet available.

Collaboration pattern

Liu works in sizeable teams: 29 distinct coauthors appear in the reviewed record, with 12 recurring collaborators — Bowen Ye, Lei Li, Shicheng Li, Zihao Yue, Linghao Zhang, Hanglong Lv, Yuanxin Liu, Wenhan Ma, Hao Tian, Rang Li, Jinhao Dong, and Yikai Zhao. Both 2026 papers place Liu in a middle authorship position, not corresponding author, consistent with a contributing role within collaborative LLM research groups rather than leading a lab.

Research keywords

reinforcement learningagentic codingRL environment synthesiscode generation agentsGRPOverifiable rewardspretraining data detectionmembership inferencelarge language modelspredictive entropyHelmholtz free energydata contamination

Highlights

CodeMidas: 5,545 RL environments synthesized from source code alone
The 2026 preprint claims that source code itself can be scaled into 5,545 reinforcement learning environments that lift MiMo-V2.5 on five benchmarks. Evolve judged the gains plausible but flagged single-run results without significance tests, assigning low reproduction readiness.
Free-energy ETD score for detecting pretraining data in LLMs
A 2026 preprint scores pretraining-data membership as model loss minus a lambda-weighted entropy term, grounded in the Helmholtz free energy view. Gains are small but consistent across 3 benchmarks; the theory is neat though the bounds look vacuous, and reproduction readiness is rated high.
A two-track LLM agenda: capability and privacy
The reviewed record splits evenly (50/50) between LLM coding agents and LLM privacy/membership inference, pairing an agentic-capability line (GRPO, verifiable rewards) with a data-privacy line (membership inference, data contamination) — a breadth that currently yields a focus score of 0.

Timeline

2026Co-authored CodeMidas (arXiv), proposing 5,545 reinforcement learning environments synthesized from source code for agentic coding, with middle authorship.
2026Co-authored the free-energy ETD paper (arXiv) on detecting pretraining data in large language models, also in a middle authorship position.
2026Both reviewed papers entered the Evolve catalog as non-corresponding work; the record spans 29 distinct coauthors across the two LLM subfields.

Papers in the Evolve catalog

Frequent co-authors

Bowen Ye×1 Lei Li×1 Shicheng Li×1 Zihao Yue×1 Linghao Zhang×1 Hanglong Lv×1 Yuanxin Liu×1 Wenhan Ma×1 Hao Tian×1 Rang Li×1 Jinhao Dong×1 Yikai Zhao×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.

Bibliometric signals (collaboration 68, focus, reproducibility 60) are derived from public bibliometric data; output, impact, and activity scores are missing entirely.
The evidence base is limited to two 2026 arXiv preprints reviewed by Evolve, both with middle authorship and neither corresponding, so the profile may not capture this researcher's full record.
'Qi Liu' is a common name and no ORCID is linked; this profile is name-resolved from public bibliographic data and may conflate same-named authors.
The reproducibility score (60) averages only 2 scored papers and should be read as early-stage evidence.