ML researcher at Amazon on efficient LLM reasoning and agent prompt optimization; first author of 2026 RIPPLE preprint.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
3
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: OpenAlex cannot confirm a clean total publication count for this name, and only 2 papers are verified in the Evolve catalog. This axis would measure cumulative output volume — papers per year and its growth over time.
Impact—
Not yet scored: the 2026 preprints are too recent for meaningful citation coverage, even though the snapshot lists an h-index of 3 and i10-index of 2. This axis would measure citation-driven influence, field-weighted reach, and uptake over time.
Focus0/100
The radar focus score is 0, reflecting a raw field distribution split evenly (0.5/0.5) between efficient reasoning in large language models and LLM agents/prompt optimization, plus OpenAlex topic hits as far afield as carbon nanotube composites and plasma diagnostics. The verified catalog is far more concentrated — both 2026 papers sit in adjacent LLM-efficiency and agent areas — so much of the apparent dispersion likely reflects same-name disambiguation noise rather than genuine scattering.
Collaboration58/100
Collaboration scores 58 on the strength of 17 distinct coauthors in the OpenAlex record. The catalog shows a strongly team-based arXiv pattern: first authorship on RIPPLE and a middle author position on RECAP. Recurring partners include Hong Wang and Jie Yuan, each listed twice among frequent coauthors, alongside Zhuoer Wang, Samson Koelle, Nikki Lijing Kuang, Yuqing Zhou, Yanjun Lin, James Feng, Ziwei Zhu, and Wei Niu.
Activity47/100
Activity scores 47, consistent with an early but currently active publishing rhythm. The counts-by-year series records 3 works in 2026 and 4 works across the recent three-year window (recent3yWorks), while sparse 2010 (2) and 2012 (1) entries almost certainly stem from same-name conflation. All verified output is 2026 arXiv preprints, so current momentum is real but the track record is short.
Reproducibility38/100
Reproducibility averages 38 across the 2 papers Evolve scored (reproMean 38, reproPapers 2). RECAP is rated medium readiness: the method is concrete — dependency-graph credit plus answer-likelihood efficacy on top of GRPO — but hinges on annotation. RIPPLE is rated low: its +23.1pp Haiku validation gain comes on a 39-task synthetic benchmark that the Evolve review flags as small-scale.
Research summary
Manqing Mao is a machine-learning researcher at Amazon in the United States whose verified recent work concentrates on making large language models reason more efficiently and on improving LLM agents through prompt optimization. The Evolve catalog contains two 2026 arXiv preprints. As first author, Mao led RIPPLE (Local Edits, Global Ripples), which gates prompt edits through advancing-policy replay and lifts Haiku validation accuracy by 23.1 percentage points on a 39-task synthetic workflow-synthesis benchmark. As a middle author, Mao contributed to RECAP (Giving Credit Where It's Due), which adds dependency-graph credit assignment and answer-likelihood efficacy to GRPO, cutting tokens by 8-31% on 7B math reasoning while improving pass@1. The OpenAlex field distribution splits evenly between these two areas (0.5 each), and the profile records 17 distinct coauthors, an h-index of 3 with an i10-index of 2, and 3 works in 2026 (4 over the recent three-year window). Recurring collaborators include Hong Wang and Jie Yuan, alongside Zhuoer Wang, Samson Koelle, and others. Radar scores place collaboration at 58, activity at 47, and reproducibility at 38, while output and impact remain unscored pending citation coverage. Because no ORCID is linked and OpenAlex topics include unrelated areas (carbon nanotube composites, plasma diagnostics), this profile is name-resolved from public bibliographic data.
Collaboration pattern
Mao's record shows a strongly collaborative arXiv pattern: 17 distinct coauthors in the OpenAlex snapshot, with Hong Wang and Jie Yuan each appearing twice among frequent partners. The 2026 catalog papers are team efforts — first authorship on RIPPLE and a middle position on RECAP — alongside recurring collaborators such as Zhuoer Wang, Samson Koelle, Nikki Lijing Kuang, Yuqing Zhou, Yanjun Lin, James Feng, Ziwei Zhu, and Wei Niu, consistent with an industrial research group at Amazon.
Research keywords
Highlights
RIPPLE: replay-gated prompt adaptation lifts Haiku validation +23.1pp
RECAP: redundancy-aware credit assignment trims 7B math tokens by 8-31%
A twin-track LLM agenda: efficient reasoning plus agent workflow synthesis
The 2026 output splits evenly across the two halves of the OpenAlex field distribution (0.5/0.5) — token-efficient chain-of-thought reasoning via RL with verifiable rewards, and persistent prompt-policy adaptation for LLM agents — suggesting a deliberate efficiency-plus-agents research program at Amazon.
Timeline
- 2026First author on RIPPLE (Local Edits, Global Ripples), an arXiv preprint on replay-informed policy adaptation for workflow synthesis, with +23.1pp Haiku validation on a 39-task synthetic benchmark.
- 2026Middle author on RECAP (Giving Credit Where It's Due), adding redundancy-aware credit assignment to GRPO for efficient reasoning, with 8-31% token reduction on 7B math.
- 2026OpenAlex counts 3 works in 2026 and 4 over the recent three-year window; both catalog papers enter Evolve review with medium (RECAP) and low (RIPPLE) reproduction readiness.
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 an OpenAlex snapshot; identity is resolved without an ORCID, so same-name conflation is possible (e.g., the 2010/2012 counts and unrelated topics such as carbon nanotube composites and plasma diagnostics).
- h-index (3) and i10-index (2) are quoted verbatim from the snapshot; only 2 papers have been reviewed in depth by Evolve.
- Both verified papers are 2026 arXiv preprints; peer-reviewed venue information is not yet available.
- Output and impact radar axes are unscored because citation coverage for the very recent preprints is insufficient.