RL researcher working on offline-RL data augmentation and online-RL behavior-prior distillation
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 output metrics are available yet: the snapshot records no total publication count or yearly paper volume beyond the two reviewed 2026 arXiv preprints. This axis would measure cumulative and per-year publication productivity across indexed venues.
Impact—
Impact data is missing: the snapshot contains no citation counts, h-index, or venue-level influence figures for this name-resolved profile. This axis would measure how widely the work is cited and its influence on offline and online reinforcement learning research.
Focus0/100
Evolve's radar records a focus score of 0, the lowest band, yet the underlying field distribution is entirely concentrated: the two reviewed papers split evenly between reinforcement learning (0.5) and machine learning with an RL label (0.5). Read together, the topical scope is narrow — offline and online RL method work — so the zero score more likely reflects sparse field metadata than demonstrated breadth.
Collaboration41/100
The collaboration score is 41, a moderate band, grounded in 7 distinct co-authors across the reviewed record. The recurring network is compact: Gong Gao, Weidong Zhao, Xiao Lai, Jiaji Shen, and Ning Jia appear as frequent collaborators. Liu holds non-leading positions — last author on the offline RL paper and middle author on the online RL paper — consistent with a contributor within a stable small team rather than a solo or large-consortium pattern.
Activity—
Activity metrics are unavailable: apart from both papers being dated 2026, the snapshot has no publication cadence or year-over-year volume data. This axis would measure recent research pace, such as papers per active year and recency of latest output.
Reproducibility38/100
Reproducibility scores 38, averaged over the 2 papers with readiness ratings. The offline RL paper (BADA) is rated medium readiness, while the online RL paper (B2PD) is rated low, and Evolve's verdicts flag thin — in B2PD's case self-contradictory — theoretical grounding in both. The sub-40 mean signals that independent replication would currently require substantial re-derivation effort; releasing code, seeds, and tightened claims would lift it.
Research summary
Xianhui Liu is an early-stage reinforcement learning researcher whose visible record centers on sample-efficiency and generalization problems across both offline and online RL. Two 2026 arXiv preprints sit in the Evolve catalog, with Liu in non-leading positions — last author on the offline RL study and a middle author on the online RL study — consistent with a contributor inside a small team of seven distinct co-authors. The first paper introduces BADA, a boundary-aware data augmentation scheme built on neighbor-bounded Mixup interpolation; Evolve judges its broad D4RL gains on scarce data credible while flagging thin theory and overreaching SOTA claims, and rates its reproducibility readiness medium. The second, B2PD, distills Q-guided CVAE behavior priors into SAC and TD3, yielding strong online-RL benchmark gains, though the theory is thin and self-contradictory and readiness is rated low. The radar shows a moderate collaboration score (41) over a compact co-author network, a focus score of 0 alongside a field distribution fully concentrated in reinforcement learning, and a reproducibility mean of 38 across two papers. Output, impact, and activity axes lack sufficient data to score. Institution, country, and ORCID are not recorded; this profile is name-resolved from public bibliographic data.
Collaboration pattern
Liu works within a compact, stable co-author network: 7 distinct collaborators appear in the record, with Gong Gao, Weidong Zhao, Xiao Lai, Jiaji Shen, and Ning Jia the most frequent. Authorship positions are non-leading — last on the offline RL paper, middle on the online RL paper, and neither corresponding — suggesting a contributing member of a small reinforcement-learning methods group rather than a principal investigator or solo author.
Research keywords
Highlights
BADA: neighbor-bounded data augmentation for offline RL
B2PD: Q-guided CVAE behavior priors distilled into SAC and TD3
Timeline
- 2026BADA — boundary-aware data augmentation for offline reinforcement learning — appears on arXiv; Liu is last author and Evolve rates reproducibility readiness medium.
- 2026B2PD — bidirectional behavior prior distillation for online reinforcement learning — appears on arXiv; Liu holds a middle author position and Evolve rates readiness low.
- 2026Evolve reviews both preprints, recording a reproducibility mean of 38 across the two papers and 7 distinct co-authors.
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.
- Only two 2026 arXiv preprints underlie this profile; no journal or conference publications are recorded in the snapshot.
- Output, impact, and activity axes have insufficient data and are left unscored on the radar.
- Institution, country, and ORCID are not recorded; the profile is name-resolved from public bibliographic data and the identity is not independently confirmed.
- Evolve verdicts and reproducibility ratings reflect editorial review of the two cataloged papers, not community citation evidence.