Jay Iyer

—

Unclaimed · AI-structured profileOpenAlex

Empirical ML-systems researcher charting LLM inference: Pareto atlases, KV-cache placement, and learned request routing

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—

No output metrics are available in the current data, so publication volume, venue mix, and productivity trends cannot be assessed. This axis would measure total papers, venues, and yearly output once a complete bibliographic record is linked.

Impact—

Impact data is not yet available: no citation counts, h-index, or venue-level influence can be quoted. This axis would measure citation uptake and scholarly influence once citation sources are linked to the author record.

Focus0/100

The focus score is 0, and the raw field distribution explains why: reviewed work splits almost evenly — about 33% each — across three overlapping labels (machine learning systems for LLM inference/serving, ML systems for efficient inference with memory hierarchy and caching, and ML systems for LLM inference serving). The spread reflects one tightly themed problem area labeled three ways rather than scattered interests; all three 2026 papers sit inside LLM inference systems.

Collaboration36/100

The collaboration score is 36, with 5 distinct coauthors across the three reviewed papers: Srikanta Datta Tumkur, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly. This points to a compact, recurring team rather than an extended network. On all three papers the author sits in a middle position and is never the corresponding author, consistent with contribution within a stable group.

Activity—

Activity cannot be scored yet, since no yearly publication timeline or recent-output data is linked to this profile. This axis would measure research cadence — publication frequency and recency — across the available record.

Reproducibility55/100

Reproducibility scores 55, the mean readiness rating across 3 evaluated papers, each of which Evolve rates 'medium' readiness. The pattern is consistent: real hardware measurements — 54 Qwen2.5-7B anchors on 3 GPUs, an 8-A40 routing testbed — alongside simulation-based results whose frontier claims await validation, and honest reporting of negative findings.

Research summary

Jay Iyer is a machine-learning-systems researcher whose reviewed work concentrates on the engineering economics of large language model inference. All three papers cataloged by Evolve appeared in 2026 on arXiv and attack LLM serving from complementary angles. 'The Inference Engineering Pareto Atlas' measures 54 Qwen2.5-7B optimization anchors on 3 GPUs to map which techniques — quantization (AWQ, FP8), KV cache compression, speculative decoding — dominate the cost-quality-latency frontier, with Evolve judging it candid and useful though simulation-dependent. 'Where Should the KV Cache Live?' examines placement policies across GPU, CPU, and SSD for long-lived sessions, finding that the reported 73× session gain is pure capacity rather than policy, while predicted-reuse and prefetch recommendations fail. 'Calibrate, Then Route' delivers a credible hardware result, showing a calibrated learned router topping goodput on 8 A40s (0.864 vs 0.835-0.847) for disaggregated serving, albeit at small scale across 3 traces without significance tests. Bibliometrically, the record is narrow but coherent: a focus score of 0 reflects an even ~33% split across three overlapping LLM-inference field labels rather than scattered interests; collaboration scores 36 with 5 distinct coauthors and middle authorship on every paper; reproducibility averages 55 ('medium' readiness on all three). Output, impact, and activity metrics are not yet available.

Collaboration pattern

Jay Iyer's reviewed work is team-authored: all three 2026 papers share a pool of just 5 distinct coauthors — Srikanta Datta Tumkur, Mehar Simhadri, Sai Pavan Kumar, Sai Kapil Kumar, and Ramesh Nampelly — and the collaboration score is 36. Holding a middle authorship position on every paper without corresponding-author credit, the pattern suggests a steady contributing role inside one stable ML-systems group rather than rotating external partnerships.

Research keywords

LLM inferenceLLM servingPareto frontier analysiscost-quality-latency tradeoffquantization (AWQ, FP8)KV cache compressionspeculative decodingmemory tieringKV cache placement policyGPU HBM offloadingprefetchingdisaggregated serving and request routing

Highlights

Mapping the inference optimization frontier across 54 anchors
Evolve's verdict on the Pareto Atlas paper: 54 Qwen2.5-7B anchors measured on 3 GPUs to simulate a cost-quality-latency atlas spanning quantization (AWQ, FP8), KV cache compression, and speculative decoding — candid and useful, though the frontier rests on unvalidated simulation.
Testing whether KV-cache tiering policy actually matters
Evolve's verdict on the KV placement study: the reported 73× session gain comes purely from added capacity, not placement policy, and the predicted-reuse and prefetch recommendations fail — honest work, but limited to a batch-one simulation.
A calibrated learned router that tops goodput on real hardware
Evolve's verdict on the routing study: a calibrated learned router leads goodput on 8 A40s (0.864 vs 0.835-0.847) for disaggregated LLM serving — a credible hardware result, though at small scale with 3 traces and no significance tests.

Timeline

2026Co-authored 'The Inference Engineering Pareto Atlas', a measurement study of 54 Qwen2.5-7B optimization anchors on 3 GPUs.
2026Co-authored the KV-cache placement study across GPU, CPU, and SSD for long-lived sessions.
2026Co-authored 'Calibrate, Then Route', the learned request routing study for disaggregated LLM serving.

Papers in the Evolve catalog

Frequent co-authors

Srikanta Datta Tumkur×3 Mehar Simhadri×3 Sai Pavan Kumar×3 Sai Kapil Kumar×3 Ramesh Nampelly×3

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.

Dimension scores come from public bibliometric data only; output, impact, and activity axes lack sufficient underlying data and are left unscored.
No institution, country, or ORCID was provided; the profile is name-resolved from public bibliographic records, so identity confidence is limited.
All three reviewed papers are 2026 arXiv preprints with middle authorship; no peer-reviewed venue, citation, or corresponding-author evidence is in scope.
Reproducibility (55) reflects Evolve readiness ratings for only 3 papers, and collaboration (36) rests on 5 coauthors — a small sample.