Salesforce researcher in efficient LLM inference and LLM agents, from KV cache eviction to multi-turn tool-use RL
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
9
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: this profile lacks a validated output index. The axis would measure publication volume and consistency over time; raw per-year work counts exist in the snapshot (1 in 2018, peaking at 28 in 2025) but remain unscored.
Impact—
Not yet scored: citation data in this snapshot is too incomplete for a reliable axis. It would measure citation volume and field-normalized influence; an h-index of 9 and i10-index of 9 are listed but are not field-normalized.
Focus0/100
Focus scores 0, reflecting minimal topical concentration: the field split is 50% machine learning and 50% the KV-cache-compression subfield, while topic clusters spread across five areas — topic modeling (36 works), NLP techniques (20), AI-based planning (17), reinforcement learning in robotics (14), and AI applications (13). The recent catalog papers narrow toward LLM inference efficiency and agentic RL, but the overall record spans multiple topics.
Collaboration61/100
Collaboration scores 61, based on 20 distinct co-authors. The recurring circle is industry-centered: Zixiang Chen, Wenting Zhao, Jianguo Zhang, Zhiwei Liu, Akshara Prabhakar, Silvio Savarese and Huan Wang appear repeatedly, consistent with a dense corporate research lab network. Both 2026 catalog papers are middle-author positions, indicating contribution within larger team-led projects rather than solo-led studies.
Activity100/100
Activity scores 100, the strongest axis: recorded outputs climb from 1 work in 2018 and 2 in 2019 to 12 in 2023, 17 in 2024, and a peak of 28 in 2025, with 14 already logged in 2026. The last three years alone account for 59 works, indicating a sustained, steeply rising publication cadence centered on LLM research.
Reproducibility78/100
Reproducibility scores 78, the mean readiness across 2 Evolve-reviewed papers. Random Attention rates high: a rigorous null result re-testing KV cache eviction across 4 models and 6 tasks. Critical-State RL rates medium: its nested-sampling diagnostic rests on four-cell BFCL experiments limited to one benchmark. The pattern suggests benchmark-driven, checkable work, though the evidence base spans only two reviewed papers.
Research summary
Shelby Heinecke is a researcher at Salesforce (United States) whose recorded work concentrates on machine learning and, increasingly, efficient LLM inference and LLM agents. The bibliometric snapshot (identity resolved via OpenAlex; ORCID 0000-0002-8831-0753) lists an h-index and i10-index of 9, and a sharply accelerating output curve: from single-digit annual works in 2018–2022 to 12 in 2023, 17 in 2024, and a peak of 28 in 2025, with 59 works in the most recent three years. Topic coverage is broad — topic modeling (36 works), NLP techniques (20), AI-based planning (17), reinforcement learning in robotics (14) — while the field split is even between general machine learning and KV cache compression for efficient inference. Two 2026 papers reviewed by Evolve capture the current agenda: Random Attention, a rigorous null result showing that prompt protection plus random per-head eviction matches the strongest KV-eviction baselines across 4 models and 6 tasks (high reproducibility readiness), and Critical-State RL, a nested Monte Carlo diagnostic that selects which tool-call state to train in multi-turn agent RL (medium readiness). Both are middle-author papers within a dense, industry-lab co-author network of 20 distinct collaborators.
Collaboration pattern
Collaboration scores 61 on 20 distinct co-authors, with a tightly recurring core: Zixiang Chen, Wenting Zhao, Zhepeng Cen, Akshara Prabhakar, Jielin Qiu, Jianguo Zhang, Zhiwei Liu, Tulika Manoj Awalgaonkar, Liangwei Yang, Silvio Savarese, Huan Wang and Heng Wang — a pattern consistent with a corporate research lab. Both reviewed 2026 papers list Heinecke as a middle author and not corresponding, suggesting contributions inside team-led LLM projects rather than solo-led studies.
Research keywords
Highlights
Random Attention: a rigorous null result on KV cache eviction
Critical-State RL: diagnosing which tool-call states to train
Publication output in steep acceleration since 2023
Recorded works climb from 12 in 2023 to 17 in 2024 and a peak of 28 in 2025, with 59 works in the most recent three years — the strongest signal in this profile's radar.
Timeline
- 2018First recorded output in the snapshot: 1 work; annual counts stay low (1–3) through 2021.
- 2023Output accelerates: 12 recorded works, up from 2 in 2022.
- 2025Peak year: 28 recorded works.
- 2026Two Evolve-reviewed papers appear — Random Attention (KV cache eviction) and Critical-State RL (multi-turn tool RL); 14 works logged so far.
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 figures are computed from OpenAlex records (identity: openalex) and may undercount or misattribute publications.
- The output and impact axes are unscored; field-normalized citation and venue-quality data were not available in this snapshot.
- The reproducibility score averages only the 2 papers Evolve has reviewed, so it is an early-stage reading.
- Profile is name-resolved from public bibliographic data (ORCID 0000-0002-8831-0753); scope is limited to the provided snapshot.