Liang Zhao (Emory University): efficient LLM inference and RL environments for coding agents
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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 output axis measures volume and consistency of publications over time. It is currently unscored: the OpenAlex identity shows only 2026 works (12 in counts-by-year) and no usable career-level output history, so this axis cannot be measured reliably yet.
Impact—
The impact axis measures citations, citation velocity, and uptake by other researchers. It is unscored here: h-index and i10-index read 0 on the OpenAlex record, indicating too little citation data to evaluate impact fairly.
Focus0/100
Focus scores 0, the lowest reading in this radar. The field distribution splits exactly 50/50 between efficient LLM inference (architecture design) and machine learning for LLM coding agents, and the OpenAlex topic list spreads across at least five clusters — topic modeling (5 works), neural network applications, NLP techniques, planning, and few-shot learning (2 each). Work is broad rather than concentrated; the reviewed 2026 papers themselves sit in the two named LLM fields.
Collaboration71/100
Collaboration scores 71, based on 34 distinct coauthors in the OpenAlex record. Twelve recurring collaborators are listed, including Hailin Zhang, Fuli Luo, Jianyu Wei, Yizhao Gao, and Yu Cheng, spanning both reviewed areas. On both reviewed 2026 papers Zhao appears as a middle-position, non-corresponding author, consistent with contributing within larger multi-author teams rather than leading small groups. The breadth of the coauthor pool suggests active cross-team work; leadership roles are not yet evidenced in the reviewed sample.
Activity75/100
Activity scores 75. The OpenAlex counts-by-year attribute 12 works to 2026 and zero to earlier years, with recent3yWorks also 12 — an entirely front-loaded, first-year burst rather than a sustained multi-year cadence. Both reviewed papers are 2026 arXiv preprints, reinforcing that the record captures a new, intensive phase. Because the history is so short, the score reflects current velocity; whether the pace holds across future years is not yet observable in this data.
Reproducibility20/100
Reproducibility scores 20 (mean 20 across 2 reviewed papers), with both rated low readiness. Evolve's verdict on CodeMidas notes that claimed gains across five benchmarks come from single runs without significance tests; HySparse2's comparisons are confounded by an MQA switch. No released code or artifacts are recorded in this data. As arXiv preprints, both may still gain released code and fuller experimental protocols, which would lift this axis in later reviews.
Research summary
Liang Zhao is an early-career researcher affiliated with Emory University (US) whose profile is name-resolved from OpenAlex public bibliographic data. The record is heavily front-loaded in 2026: OpenAlex attributes 12 works to this identity in 2026 (recent3yWorks = 12), and Evolve has reviewed two of these arXiv papers. The work splits evenly (0.5/0.5) between efficient large language model inference — architecture design such as sparse attention and KV-cache sharing (HySparse2, 2026) — and machine learning for LLM coding agents, specifically reinforcement-learning environment synthesis (CodeMidas, 2026). On both reviewed papers Zhao is a middle author and not corresponding, consistent with a contributor role inside larger teams. Collaboration is broad: 34 distinct coauthors, with 12 recurring collaborators including Hailin Zhang, Fuli Luo, and Jianyu Wei (collaboration score 71). Activity scores 75 on the strength of the 2026 output burst, but the identity's citation footprint is not yet usable (h-index 0, i10-index 0 on OpenAlex), so output and impact are left unscored. Reproducibility is low (mean 20 across the 2 reviewed papers): Evolve flags single-run RL results without significance tests in CodeMidas and a confounded MQA switch in HySparse2 comparisons. Topic labels on the OpenAlex record (e.g., topic modeling, 5 works) look broader than the reviewed output and may reflect name-collision noise.
Collaboration pattern
Liang Zhao's record lists 34 distinct coauthors and 12 recurring collaborators, including Hailin Zhang, Fuli Luo, Jianyu Wei, Yizhao Gao, Qihao Zhang, and Yu Cheng. In both reviewed 2026 papers Zhao sits in a middle author position without corresponding authorship, suggesting contribution within larger lab or industry-style teams rather than solo leadership. The coauthor pool spans both active areas — efficient inference and coding-agent RL — indicating parallel collaborations across the two lines of work.
Research keywords
Highlights
Hybrid sparse attention with two-level KV sharing
RL environments synthesized from code itself
A 2026-only, evenly split output record
OpenAlex attributes 12 works to this identity, all in 2026, and the field distribution splits 50/50 between efficient LLM inference (architecture design) and machine learning for LLM coding agents — a new, two-track portfolio spanning systems work and RL for coding agents.
Timeline
- 2026HySparse2 appears on arXiv: a middle-author paper on hybrid sparse attention with two-level KV sharing for long-context LLM inference.
- 2026CodeMidas appears on arXiv: synthesizing 5,545 reinforcement-learning environments from source code to lift MiMo-V2.5 on five benchmarks.
- 2026OpenAlex attributes 12 works to this identity in 2026 (recent3yWorks: 12), marking the visible start of the record's output curve.
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.
- Bibliometric scores are computed from OpenAlex data only; no ORCID link was available to verify identity.
- "Liang Zhao" is a common name: this profile is name-resolved from public bibliographic data and may merge or miss records.
- Citation metrics (h-index 0, i10-index 0) are not usable, so the output and impact dimensions are left unscored.
- Only 2 Evolve-reviewed papers (both 2026 arXiv preprints) underlie the reproducibility score and highlights; the other 2026 works are not individually reviewed.