Salesforce researcher bridging statistical generative modeling and long-horizon LLM agent architectures
2
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
17
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Publication volume metrics are not available in this snapshot, so this axis is unscored. It would normally aggregate total publication count, per-year productivity, and authorship-position distribution to characterize overall research throughput across the career.
Impact—
Citation-based impact data are insufficient in the current snapshot, leaving this axis unscored. It would measure total citations, citation velocity, and highly cited works — beyond the h-index (17) and i10-index (18) already on file.
Focus100/100
A focus score of 100 indicates maximal topical concentration in this snapshot. Bibliometric topic assignments cluster tightly in statistical machine learning — topic modeling (19), generative adversarial networks and image synthesis (13), Markov chains and Monte Carlo methods (12) — and the recent field distribution is tagged entirely as AI agents (LLM systems). Both Evolve-reviewed 2026 papers sit in the same long-horizon agent space, reinforcing a single coherent line.
Collaboration28/100
A collaboration score of 28 sits low, consistent with a compact network: only 3 distinct coauthors appear in the recent window, with Egor Pakhomov, Anurag Koul, and Bo Pang as recurring names. Both catalog papers follow a small-team pattern — one first-authored, one last-authored — rather than large consortium work, suggesting deep, stable partnerships over broad collaboration.
Activity61/100
An activity score of 61 reflects a moderate, uneven output rhythm. Annual counts peaked at 10 works in 2020 and 9 in 2019, eased to 5 in 2021 and 3 in 2022, rose again to 7 in 2023, and stand at 2 works in 2025 with 5 logged for 2026. The last three years contribute 7 works, indicating sustained but selective current engagement.
Reproducibility20/100
Reproducibility evidence is limited: the mean readiness score is 20, drawn from just one evaluated paper. Evolve's verdicts flag concrete constraints — the agent-architecture paper reports a single 10-day autonomous campaign with no baselines (low readiness), and the trace-model paper's benchmarks were co-developed with the method (medium readiness). Strong controlled numbers (30/30 vs 8/30) exist, but independent replication is not yet established.
Research summary
Erik Nijkamp is a researcher at Salesforce (United States); this profile is name-resolved from public OpenAlex bibliographic data. The record shows an h-index of 17 and an i10-index of 18, with annual output peaking at 9 works in 2019 and 10 in 2020, dipping through 2021-2022, recovering to 7 in 2023, and continuing with 2 works in 2025 and 5 logged for 2026. Topic assignments concentrate in statistical machine learning: topic modeling (19), generative adversarial networks and image synthesis (13), Markov chains and Monte Carlo methods (12), natural language processing techniques (9), and adversarial robustness (8). The Evolve-reviewed catalog adds a current focus on LLM-based agents: a first-author 2026 arXiv paper on a levels/ticks/cascaded-intelligence architecture for long-horizon agents, and a last-author 2026 paper on a live trace model for agent observability. Recent output is moderate (7 works in three years) within a small coauthor set (3 distinct collaborators), matching a low collaboration score of 28. Focus is scored at the maximum of 100, while reproducibility evidence is thin (mean 20 from one evaluated paper).
Collaboration pattern
The collaboration footprint is narrow: 3 distinct coauthors in the recent window and a collaboration score of 28. Egor Pakhomov, Anurag Koul, and Bo Pang recur as frequent coauthors, pointing to a small, stable working group. Both 2026 catalog papers are small-team efforts — one first-authored, one last-authored — consistent with deep bilateral collaborations rather than large multi-institution projects.
Research keywords
Highlights
A levels/ticks/cascade architecture for long-horizon LLM agents
One trace stream serving both agent and observer
Statistical ML foundations beneath the agent work
Bibliometric topic assignments show depth outside the agent space: topic modeling (19 works), generative adversarial networks and image synthesis (13), and Markov chains and Monte Carlo methods (12) — a generative-modeling core that predates the current LLM-agent focus.
Timeline
- 2026Publishes a first-author arXiv paper on a levels/ticks/cascaded-intelligence architecture for long-horizon agents, demonstrated with a 10-day autonomous RL reproduction campaign.
- 2026Publishes a last-author arXiv paper on a live trace model for long-horizon agents, with 30/30 versus 8/30 controlled comparisons at 14x token cost.
- 2026Evolve review assigns repro-readiness ratings of low and medium to the two papers, citing single-campaign evidence and co-developed benchmarks.
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.
- Identity is name-resolved from a public OpenAlex record; no ORCID is linked, so the bibliometric snapshot may merge or split same-name authors.
- Bibliometric indicators come from OpenAlex only; output and impact dimension scores are missing from this snapshot.
- Qualitative assessments rest on two 2026 arXiv preprints in the Evolve catalog; the reproducibility mean is based on a single evaluated paper.
- The recent field distribution covers only one tagged work, so field-level classification rests on thin coverage.