Amazon researcher: efficient reasoning in LLMs and LLM-agent prompt optimization — credit assignment and replay gating
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
Output is unscored: the snapshot does not provide a complete indexed publication record for this author, so total volume cannot be computed. This axis would measure cumulative works, venue spread, and per-year output distribution.
Impact—
Impact is unscored: citation counts, h-index, and i10-index are not populated for this author in the current snapshot, so influence cannot be estimated. This axis would measure citation uptake, citing works, and per-paper attention.
Focus0/100
The focus score is 0, reflecting a perfectly even 0.5/0.5 field distribution split between 'machine learning — efficient reasoning in large language models' and 'machine learning — LLM agents and prompt optimization'. Topic records — topic modeling (3), NLP techniques (2), AI-based problem solving and planning (2), plus single records in information retrieval/search behavior and reinforcement learning in robotics — confirm breadth across adjacent LLM-facing areas rather than concentration in one line of work.
Collaboration58/100
Collaboration scores 58, computed from 17 distinct coauthors in the OpenAlex record. Both Evolve-reviewed papers place Koelle as a middle-position, non-corresponding author — a contributor within a larger team rather than a solo lead. Recurring partners include Manqing Mao, Zhuoer Wang, Nikki Lijing Kuang, Yuqing Zhou, Hong Wang, and Jie Yuan, with Hong Wang and Jie Yuan each appearing twice in the frequent-coauthor list, indicating a stable core group behind both 2026 papers.
Activity47/100
Activity scores 47 on a thin but current record: the counts-by-year series lists only 2026, with 4 indexed works, and recent3yWorks is likewise 4 — so essentially all indexed output falls in the current year. With no earlier-year records in the snapshot window, this score reflects recency and low accumulated volume rather than sustained multi-year productivity.
Reproducibility38/100
Reproducibility scores 38 (reproMean) across 2 Evolve-assessed papers. The RECAP paper is rated medium: its 8–31% token reduction with pass@1 gains on 7B math is checkable but hinges on annotation quality. RIPPLE is rated low: its +23.1pp Haiku validation gain rests on a small 39-task synthetic benchmark. Both works are arXiv preprints, which further limits independent verification.
Research summary
Samson Koelle is a United States–based researcher at Amazon whose indexed work centers on applied machine learning for large language models. The snapshot's OpenAlex record (name-resolved; no ORCID linked) splits the field distribution evenly at 0.5/0.5 between efficient reasoning in large language models and LLM agents with prompt optimization, with topic records led by topic modeling (3), NLP techniques (2), and AI-based problem solving and planning (2), plus single records in information retrieval/search behavior and reinforcement learning in robotics. Evolve reviews two 2026 arXiv papers, both with Koelle as a middle-position, non-corresponding author. RECAP adds LLM dependency-graph credit assignment and answer-likelihood efficacy to GRPO, cutting tokens 8–31% on 7B math while improving pass@1, though it hinges on annotation quality. RIPPLE gates prompt edits through advancing-policy replay, lifting Haiku validation +23.1pp on a 39-task synthetic benchmark. Collaboration records list 17 distinct coauthors, with recurring partners such as Manqing Mao, Zhuoer Wang, and Jie Yuan. Recorded activity is concentrated in 2026, with 4 works in the last three years; citation-based output and impact metrics are not yet available in this snapshot.
Collaboration pattern
The record lists 17 distinct coauthors (collaboration score 58), and both reviewed papers place Koelle in a middle, non-corresponding author position — a team contributor rather than a solo lead. The frequent-coauthor list is tight and overlapping, with Hong Wang and Jie Yuan each appearing twice, alongside Manqing Mao, Zhuoer Wang, Nikki Lijing Kuang, Yuqing Zhou, Yanjun Lin, James Feng, Ziwei Zhu, and Wei Niu — a stable recurring group carrying both the reasoning-efficiency and agent-adaptation lines in 2026.
Research keywords
Highlights
RECAP: redundancy-aware credit assignment for efficient LLM reasoning
RIPPLE: replay-gated prompt adaptation for workflow synthesis
A complementary 2026 method pair across two LLM subfields
The two reviewed papers attack LLM reliability from both sides — credit-aware RL training for shorter reasoning chains (RECAP), and replay-gated adaptation for safer prompt/workflow edits (RIPPLE) — mirroring the 0.5/0.5 field split in the bibliometric record.
Timeline
- 2026Co-authored 'Giving Credit Where It's Due: Redundancy-Aware Learning for Efficient Reasoning' (RECAP) on arXiv, as a middle-position, non-corresponding author.
- 2026Co-authored 'Local Edits, Global Ripples: Replay-Informed Policy Adaptation for Workflow Synthesis' (RIPPLE) on arXiv, again as a middle-position, non-corresponding author.
- 2026OpenAlex counts-by-year lists 4 indexed works for 2026 (recent3yWorks = 4) — the entirety of the current snapshot window, marking the start of the visible publication record.
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 come from a name-resolved OpenAlex author record; no ORCID is linked in the snapshot, so identity matching carries some disambiguation risk.
- Citation-based output and impact metrics (citation counts, h-index, i10-index) are not populated in this snapshot, so the output and impact radar axes are unscored.
- Both reviewed papers are 2026 arXiv preprints with no peer-reviewed venue information in the provided data.
- Reproducibility rests on only 2 Evolve-assessed papers (reproMean 38) and activity on a single-year counts-by-year record, so both scores are low-confidence.