ML researcher applying RL with verifiable rewards to constrained music generation and LLM tool selection for genomics.
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: total publication volume, venue quality, and output trajectory are unavailable in the current extract. Only 8 works in the trailing three years are recorded — too thin a base to characterize output.
Impact—
Not yet scored: the recorded h-index and i10-index are both 0, so citation-based impact cannot be measured yet. This axis would track citation counts and uptake of the 2026 papers once citations accrue.
Focus0/100
The focus score is 0, reflecting an even split: the field distribution assigns exactly 0.5 to RLVR for music generation and 0.5 to RL for LLM tool selection in genomics, while topic assignments scatter across five clusters — AI-based problem solving and planning (3 works), multimodal machine learning applications (3), topic modeling (2), reinforcement learning in robotics (2), and advanced neural network applications (1).
Collaboration46/100
The collaboration score is 46. The record lists 9 distinct coauthors, and five names recur — Ye Chen, Zhichao Wang, Xiaoyu Ma, Haoran Shou, and Xiaoying Tang — pointing to a small stable core team. On both reviewed papers Liu is first author but not corresponding author, indicating a technical-lead role within a supervised group rather than an independent research lead.
Activity64/100
The activity score is 64, based on 8 works in the trailing three years. Notably, all 8 are logged in 2026, meaning the entire visible record lands in a single year — a burst of first-author arXiv output rather than a multi-year cadence. The single-year concentration makes the trajectory hard to extrapolate but signals strong current momentum.
Reproducibility78/100
The reproducibility score is 78, computed across 2 reviewed papers. The music-generation paper rates high — Evolve credits a strong benchmark and honest statistics — while the FGPO genomics paper rates medium. Both disclose explicit, checkable numbers (0.160 to 0.807, 16 tool subsets, 15 settings), a solid reproducibility signal for preprint-stage work.
Research summary
Haoyue Liu is an early-career machine learning researcher whose record, per OpenAlex, lists an affiliation with the Chinese University of Hong Kong, working at the intersection of reinforcement learning with verifiable rewards (RLVR) and large language model agents. The record shows 8 works in the trailing three years, all logged in 2026, with h-index and i10-index both at 0 — a citation footprint consistent with a researcher just starting to publish. Topic assignments cluster around AI-based problem solving and planning (3 works) and multimodal machine learning applications (3), with smaller clusters in topic modeling, reinforcement learning in robotics, and advanced neural network applications. Evolve has reviewed two first-author 2026 arXiv papers. The first introduces graded verifier rewards for constraint-following symbolic music generation, lifting a Qwen3-4B model from 0.160 to 0.807 on constrained ABC-notation benchmarks, with high reproducibility readiness. The second proposes FGPO, replacing GRPO's sampled rollouts with exact enumeration over 16 tool subsets for genomic tool selection, winning all 15 evaluated genomic settings, though limited to tiny action spaces. The portfolio splits evenly (0.5/0.5) between music-generation RLVR and RL for LLM tool selection in genomics, with a reproducibility mean of 78 across the two reviewed papers. Nine distinct coauthors appear, including recurring collaborators Ye Chen, Zhichao Wang, Xiaoyu Ma, Haoran Shou, and Xiaoying Tang.
Collaboration pattern
Liu's record shows 9 distinct coauthors and a stable recurring core — Ye Chen, Zhichao Wang, Xiaoyu Ma, Haoran Shou, and Xiaoying Tang appear repeatedly in the frequent-collaborator list. Both catalog papers are first-authored but not corresponding-authored, consistent with carrying the technical lead within a supervised group rather than running an independent lab. The collaboration score of 46 sits mid-range: real, repeated co-authorship within a small network concentrated in one or two teams.
Research keywords
Highlights
Graded verifier rewards lift constrained music generation from 0.160 to 0.807
FGPO: exact enumeration beats sampled rollouts for genomic tool selection
A verifier-centric RLVR agenda spanning music and genomics
The two reviewed papers share one methodological stance — make the reward checkable. The field distribution splits 0.5/0.5 between music-generation RLVR and RL for LLM tool selection in genomics, with a reproducibility mean of 78 across both, suggesting a coherent early research identity built around verifier design and exact or graded reward signals.
Timeline
- 2026Published first-author arXiv paper introducing graded verifier rewards for constraint-following music generation (Evolve reproducibility readiness: high).
- 2026Published first-author arXiv paper proposing FGPO, exact policy optimization for genomic tool selection (Evolve reproducibility readiness: medium).
- 2026OpenAlex logs 8 works in the trailing three-year window, all in this single year, spanning two applied RL directions with 9 distinct coauthors.
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 bibliometrics come from a single OpenAlex extract; recorded h-index and i10-index are both 0, so no citation impact is yet measurable.
- Identity is name-resolved from public bibliographic data: 'Haoyue Liu' is a common name, and the OpenAlex country field (US) conflicts with the listed institution (Chinese University of Hong Kong), so identity confidence is limited.
- Qualitative verdicts rest on only two Evolve-reviewed papers (both 2026 arXiv preprints); the 8-work count includes records not reviewed here, and preprints are not yet peer-reviewed.
- The reproducibility score (mean 78) covers only the 2 reviewed papers; the focus score of 0 reflects the even 0.5/0.5 field split and may shift as the portfolio develops.