Salesforce researcher on LLM post-training, agentic memory, and agent evaluation benchmarks
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
19
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 is not yet scored: the underlying data on this researcher's complete output is insufficient for a reliable rating. It would measure total publication count, authorship positions, and year-by-year production volume drawn from verified sources.
Impact—
Impact is not yet scored because citation-side data for this identity is incomplete; although the snapshot lists an h-index of 19 and an i10-index of 24, Evolve has not computed the axis. It would measure citations and downstream scholarly influence.
Focus0/100
Focus scores 0, driven by an almost perfectly even split of the field distribution: machine learning (LLM post-training), LLM agent evaluation / machine learning, and machine learning — LLM-based agents and agent memory systems each account for 0.33. The three labels overlap heavily in substance — all three reviewed 2026 papers target LLM agents — so the low score reflects evenly balanced field tags rather than a scattered agenda.
Collaboration55/100
Collaboration scores 55, computed from 15 distinct co-authors in the record. Recent output is team-based: all three reviewed 2026 papers list Yavuz in a middle, non-corresponding position. Recurring co-authors include Yang Li, Shafiq Joty, Zixuan Ke, Vaidehi Patil, Haizhou Shi, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Mohit Bansal, and Yefan Zhou — a mix consistent with an industrial lab working alongside academic partners.
Activity100/100
Activity is the strongest axis at 100. OpenAlex records 43 works in the last three years, and the yearly series climbs from 2 (2013) through 9 (2019) and 11 (2022) to a peak of 20 (2024), continuing at 10 (2025) and 13 (2026). This indicates sustained and currently accelerating publication throughput.
Reproducibility55/100
Reproducibility scores 55 — the mean readiness across the 3 papers Evolve has reviewed, each rated medium. Verdicts note good ablations but no significance tests for Just-in-Time Memory, consistent gains with honest ablations but shrinkage at scale for RISE, and solid design with mechanism analysis but thin data statistics for EvoHarnessBench. Code or artifact availability is not recorded.
Research summary
Semih Yavuz is a researcher at Salesforce (United States) working on large language model post-training, LLM-based agents, and agent evaluation. The profile is name-resolved from OpenAlex bibliographic data, with no ORCID linked. The snapshot lists an h-index of 19 and an i10-index of 24, with 43 works in the last three years; yearly output climbed from 8 works in 2023 to a peak of 20 in 2024, continuing at 13 works in 2026. Topic coverage is broad — topic modeling (69 works), NLP techniques (51), speech and dialogue systems (21), information retrieval and search behavior (18), and multimodal machine learning applications (18) — while the field distribution now splits evenly across three closely related areas: LLM post-training, LLM agent evaluation, and LLM-based agents and agent memory systems. Evolve has reviewed three 2026 arXiv preprints on which Yavuz is a middle-position, non-corresponding author: Just-in-Time Memory (read-time memory curation trained with GRPO), RISE (self-extrapolating policy distillation from an RLVR trajectory), and EvoHarnessBench (a benchmark for agent adaptation to evolving harnesses). Reproducibility readiness is rated medium for all three: ablation quality is praised, while missing significance tests and thin data statistics are recurring caveats. Collaboration is team-scale, with 15 distinct co-authors in the record.
Collaboration pattern
Yavuz's collaboration pattern is team-scale industrial research: 15 distinct co-authors appear in the record, and all three reviewed 2026 papers place Yavuz in a middle, non-corresponding position. The recurring collaborator list — Yang Li, Shafiq Joty, Zixuan Ke, Vaidehi Patil, Haizhou Shi, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Mohit Bansal, Yefan Zhou — spans agent memory, post-training, and agent evaluation.
Research keywords
Highlights
Just-in-Time Memory: curation moved to read time
RISE: self-extrapolating policy distillation
EvoHarnessBench: testing agents against an evolving harness
Timeline
- 2013Earliest works in the snapshot: OpenAlex records 2 publications in 2013, the start of the indexed record.
- 2019Output reaches 9 works, an early peak in the profile's growth phase after 4 works in 2018.
- 202211 works recorded, continuing steady growth as topic coverage spans topic modeling (69 cumulative) and NLP techniques (51 cumulative).
- 2024Peak annual output: 20 works recorded in 2024, the highest single-year count in the snapshot.
- 202613 works recorded so far; Evolve reviews three arXiv preprints — Just-in-Time Memory, RISE, and EvoHarnessBench — with Yavuz as a middle-position author.
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 OpenAlex only; the profile is name-resolved from public bibliographic data and no ORCID is linked, so identity confidence is limited for a relatively common name.
- Output and impact dimensions are unscored; the h-index (19) and i10-index (24) are raw snapshot values, not Evolve-verified metrics.
- All three reviewed papers are 2026 arXiv preprints with middle authorship positions; peer-review status and code availability are not in the data.
- Counts-by-year and topic counts reflect the indexed OpenAlex record only and may undercount non-indexed output; affiliation data is limited to Salesforce (United States).