Amazon researcher advancing efficient LLM reasoning and replay-gated prompt optimization for LLM agents
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
18
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—
This axis measures publication volume and consistency, but Evolve's output score is not yet computed for this profile: per-work metadata is incomplete in the OpenAlex record, so the axis stays unscored rather than being estimated from partial counts.
Impact—
Impact scoring needs citation counts per paper, which the underlying record does not yet provide; the axis would measure citations, citation velocity and research uptake. It is left unscored here rather than inferred from the profile-level h-index alone.
Focus0/100
The focus axis scores 0 because Evolve's field distribution is split exactly evenly: 50% of reviewed output in efficient reasoning for large language models and 50% in LLM agents and prompt optimization. The broader topic record adds an unrelated cluster — supramolecular chemistry, self-assembly and crystallography topics (28, 27 and 17 works) — so no single field currently dominates.
Collaboration58/100
The collaboration axis scores 58 from 17 distinct coauthors in the record. Recurring names span Manqing Mao, Samson Koelle, Nikki Lijing Kuang, Yuqing Zhou, Hong Wang, Jie Yuan, Yanjun Lin, James Feng, Ziwei Zhu and Wei Niu. Both cataloged 2026 papers list this researcher as a middle author rather than corresponding author, consistent with team-based rather than solo-led work.
Activity100/100
The activity axis scores 100. The year-by-year record shows continuous output from 2018 (1 work) through 2026 (11 works to date), with 14 works in 2023 and 13 in 2024, and the most recent three years accounting for 34 works. Current publishing intensity sits at the top of the measured range rather than tapering.
Reproducibility38/100
The reproducibility axis scores 38, averaged over the 2 papers Evolve has reviewed. RECAP's efficient-reasoning results are rated medium readiness but hinge on annotation pipelines; RIPPLE's replay-gated prompt adaptation is rated low, with gains shown on a 39-task synthetic benchmark and Haiku validation only. Replication currently requires care.
Research summary
Zhuoer Wang is a researcher at Amazon (United States) whose reviewed work concentrates on making large language models reason and act more efficiently. Evolve's catalog includes two 2026 arXiv papers with Wang as a middle author: RECAP, which adds LLM dependency-graph credit plus answer-likelihood efficacy to GRPO for efficient reasoning (8–31% token reduction on 7B math with pass@1 gains), and RIPPLE, which gates prompt edits through advancing-policy replay for workflow synthesis (+23.1pp Haiku validation on a 39-task synthetic benchmark). The bibliometric record, name-resolved from OpenAlex, shows 34 works in the last three years, continuous annual output from 2018 through 2026 (peaking at 14 works in 2023), 17 distinct coauthors, and profile-level h-index 18 with i10-index 32. Evolve's field distribution splits evenly between efficient LLM reasoning and LLM-agent prompt optimization, which keeps the focus axis at 0, and the reproducibility axis averages 38 across the two reviewed papers (one medium, one low readiness). Notably, the topic record mixes in an unrelated supramolecular-chemistry cluster, so identity resolution should be treated cautiously for a relatively common name. Output and impact axes remain unscored pending per-paper citation metadata.
Collaboration pattern
Seventeen distinct coauthors appear in the record; the most frequent include Manqing Mao, Samson Koelle, Nikki Lijing Kuang, Yuqing Zhou, Hong Wang, Jie Yuan, Yanjun Lin, James Feng, Ziwei Zhu and Wei Niu (Hong Wang and Jie Yuan repeat in the source list). Both cataloged 2026 papers position Wang as a middle, non-corresponding author, indicating contribution within larger ML teams rather than sole leadership.
Research keywords
Highlights
RECAP: redundancy-aware credit assignment cuts math reasoning tokens by 8–31%
RIPPLE: replay-gated prompt edits lift workflow-synthesis validation by 23.1pp
Timeline
- 2018First works appear in the bibliometric record (1 publication).
- 2020Single-year output reaches 10 works.
- 2023Most productive year in the record: 14 works.
- 202413 works published, sustaining the high-output pace of 2023.
- 202611 works to date; both Evolve-reviewed papers appear on arXiv — RECAP on efficient reasoning and RIPPLE on LLM-agent workflow synthesis.
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.
- Bibliometrics are name-resolved from OpenAlex only; "Zhuoer Wang" is a relatively common name and the topic record mixes an unrelated supramolecular-chemistry cluster, so identity resolution carries uncertainty.
- Output and impact radar axes are unscored because per-paper citation metadata is insufficient, even though profile-level h-index (18) and i10-index (32) are present.
- Only 2 papers carry Evolve-reviewed verdicts, and both are 2026 arXiv preprints, so the reviewed findings are not yet peer-reviewed.
- The reproducibility score averages just 2 papers (one medium, one low readiness), a thin base for generalizing.