Jerry Kaplan

Center for International Policy · United States

Unclaimed · AI-structured profileOpenAlex

First-author researcher on calibrated LLM-judge evaluation and efficient inference — early-stage profile

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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—

The output axis (publication volume and productivity) is not yet scored: the OpenAlex-indexed record lists an h-index of 0, an i10-index of 0, and only two works, so the underlying bibliographic record is too sparse to characterize output.

Impact—

The impact axis (citations and h-index-based scholarly influence) is not scored: with an h-index of 0 and i10-index of 0 on the current OpenAlex record, there is not yet enough citation data to assess influence.

Focus0/100

The focus score is 0, and the raw field distribution splits evenly (0.5/0.5) between two adjacent labels, both centered on efficient LLM inference; topic counts (Topic Modeling, Advanced Neural Network Applications, Natural Language Processing Techniques: 2 each) suggest a tightly scoped agenda rather than dispersion, though two works are too few to settle the question.

Collaboration0/100

The collaboration score is 0 and the record lists 0 distinct co-authors: both cataloged papers are first-authored and corresponding with no frequent co-authors recorded, indicating solo work in this indexed snapshot; co-authorship patterns may change as more of the record is linked.

Activity32/100

The activity score is 32 on a thin base: the record shows 2 works in the recent three-year window, both dated 2026 (countsByYear lists only 2026 with 2 outputs). Productivity is low in absolute volume but concentrated in a single, current year.

Reproducibility100/100

The reproducibility score is 100 (reproMean 100 across 2 scored papers): both cataloged works carry a high reproducibility-readiness rating, consistent with the calibrated-judge protocol with pre-registered equivalence bounds in one paper and explicit threshold-calibration procedures in the other, making transparent methodology this profile's most distinctive measurable strength.

Research summary

Jerry Kaplan (Center for International Policy, US) is profiled here from public OpenAlex bibliographic data and two arXiv papers reviewed in the Evolve catalog; both are 2026 first-author, corresponding-author works in efficient large language model inference and its evaluation methodology. The first, "A Calibrated Instrument for Measuring How Inference Optimizations Affect Output Quality," proposes a formally calibrated LLM-judge protocol — dual-reference null, implementation null, positive controls, and a pre-registered ±0.3 equivalence bound — for comparing lossy optimizations such as quantization, speculative decoding, and early exit; Evolve's verdict praises its controlled design while noting limits to two small models, one judge family, and narrow intervention coverage. The second, "One Size Does Not Fit All," shows early-exit savings vary sharply by traffic, with per-deployment threshold gains reaching 59 points, though validation relies on a fidelity metric the companion work shows unreliable. The indexed record is thin: h-index 0, i10-index 0, 2 works in the recent three-year window, 0 distinct co-authors, and output and impact dimensions unscored. Reproducibility is the standout signal, with a perfect 100 across both papers. Given a common name and no ORCID, identity is resolved from bibliographic data, making this an early-stage profile.

Collaboration pattern

The indexed record shows 0 distinct co-authors and an empty frequent-co-author list; both cataloged papers are single-profile first and corresponding authorship. The collaboration dimension scores 0 on this basis. This suggests independently produced work in the current snapshot, though the thin OpenAlex record (2 works) means little can be concluded about longer-term collaboration habits.

Research keywords

LLM-as-a-judge calibrationquantizationspeculative decodingearly exitequivalence testingperplexity proxy failureLLM inference accelerationconfidence gatingtuned lensthreshold calibrationedge deployment

Highlights

A calibrated judge protocol puts lossy inference optimizations to a formal test
The 2026 first-author, corresponding-author paper proposes a formally calibrated LLM-judge protocol — dual-reference null, implementation null, positive controls, and a pre-registered ±0.3 equivalence bound — to compare lossy inference optimizations; Evolve credits its careful controls while flagging two small models, one judge family, and narrow intervention coverage as limits.
Early-exit depth tuned per deployment, with gains up to 59 points
The second 2026 first-author, corresponding-author paper shows early-exit savings vary sharply by traffic, with per-deployment threshold gains reaching 59 points; Evolve notes validation uses the fidelity metric the companion paper shows unreliable.
Both papers rated high reproducibility (reproMean 100)
Evolve scores reproducibility readiness at 100 across the 2 cataloged papers, both rated high — consistent with pre-registered equivalence bounds and explicit threshold-calibration procedures, the strongest measurable signal in this early-stage record.

Timeline

2026Published "A Calibrated Instrument for Measuring How Inference Optimizations Affect Output Quality" (arXiv) as first and corresponding author.
2026Published "One Size Does Not Fit All: Setting Inference Depth from the Questions a Deployment Actually Asks" (arXiv) as first and corresponding author.
2026Both cataloged papers received high reproducibility-readiness ratings (reproMean 100 across 2 papers).
2026These two works constitute the entire recent three-year record (recent3yWorks: 2; countsByYear lists only 2026).

Papers in the Evolve catalog

2026-09-17first authorcorresponding
A Calibrated Instrument for Measuring How Inference Optimizations Affect Output Quality
2026-09-12first authorcorresponding
One Size Does Not Fit All: Setting Inference Depth from the Questions a Deployment Actually Asks

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 single OpenAlex record; no ORCID was provided, and "Jerry Kaplan" is a common name, so this profile is name-resolved from public bibliographic data.
Output and impact dimensions are unscored because the indexed citation record (h-index 0, i10-index 0, 2 works) is too sparse.
The profile rests on 2 arXiv papers reviewed by Evolve, both from 2026; this snapshot contains no information on earlier or non-arXiv work.
Reproducibility scores reflect Evolve's reproducibility-readiness assessment (2 papers), not external replication studies.