Shekar Ramachandran

—

Unclaimed · AI-structured profileOpenAlex

Efficient-LLM researcher compressing transformer depth and quantization while stress-testing persistent-memory agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

The output axis is unscored: the snapshot lacks a verified total-works count, so publication volume and consistency across the full career cannot be computed. It would measure cumulative output volume, venue spread, and year-over-year consistency.

Impact—

Impact is unscored. The snapshot reports an h-index of 1 and i10-index of 1, too thin a citation base to rank. This axis would measure citation uptake, citing works, and influence on follow-up research.

Focus—

Focus is unscored because topic classification covers too little of the record to compute concentration. It would measure how tightly output clusters; here, Topic Modeling (5) and NLP techniques (3) hint at an inference-and-agents center of gravity.

Collaboration14/100

The collaboration axis scores 14/100, the weaker of the two computed dimensions. The OpenAlex snapshot records only 1 distinct coauthor, with Jundong Hu as the recurring partner, and all three 2026 Evolve-catalog papers place this researcher in the last author position without corresponding-author designation. The pattern suggests a compact, stable two-person working group rather than a broad consortium.

Activity61/100

Scoring 61/100, activity is this profile's stronger axis. The snapshot counts 7 works in the recent three-year window, with yearly output rising from a single 2023 work to 7 works in 2026. All three Evolve-reviewed papers appeared on arXiv in 2026, indicating a concentrated, currently active phase rather than a long publication history.

Reproducibility—

Reproducibility is unscored at the profile level. It would aggregate repro-readiness signals such as code release, seeds, and statistical testing across all outputs. The three reviewed papers carry readiness labels of medium, high, and medium, but coverage is too sparse to score.

Research summary

Shekar Ramachandran is a researcher whose public record suggests an early-career stage, working on efficient large language model inference and the reliability of LLM agents. The OpenAlex-based snapshot records an h-index of 1 and i10-index of 1, one work in 2023, and a step up to seven works in 2026, giving a recent three-year total of seven works and an activity score of 61/100. Topic assignments concentrate on Topic Modeling (5), NLP techniques (3), and advanced neural network applications (2), with single appearances in software engineering research and information retrieval. The Evolve catalog reviews three 2026 arXiv papers, all with this researcher in the last author position: XMerge on cross-axis selection and reconstructive layer merging for LLM depth compression; a causal study showing quantization damage in LLMs is diffuse and best spent on global g128 granularity rather than oracle local repair; and The Memory Trust Gap, which argues stale-memory over-trust in persistent-memory agents is capability-gated. Evolve's verdicts praise well-controlled designs while flagging single-seed runs, synthetic-task evidence on 0.6-8B models, and partially separable statistical margins. Collaboration is narrow: one distinct coauthor, with Jundong Hu the recurring partner, yielding a collaboration score of 14/100.

Collaboration pattern

Collaboration is the profile's narrowest dimension, scoring 14/100. The OpenAlex snapshot lists exactly 1 distinct coauthor, and Jundong Hu appears as the frequent collaborator, indicating a stable two-person working pattern. All three 2026 Evolve-catalog papers place this researcher in the last author slot, though none designates them as corresponding author. Expanding the co-author network or confirming corresponding-author roles would strengthen collaborative-visibility signals.

Research keywords

LLM depth compressionLayer pruning and mergingPost-training quantizationMixed precisionEfficient LLM inferencePersistent-memory agentsStale information over-trustModel scalingTopic modelingNatural language processing

Highlights

XMerge: select-then-refit layer merging for LLM depth compression
Evolve's review credits the cross-axis selection and reconstructive merging design with broad results across 7 models, while noting a single seed per setting and statistically separable CORE margins in only 3 of 7 cases; reproducibility readiness is rated medium.
The memory trust gap: stale-memory over-trust is capability-gated
The reviewed paper claims stale-memory over-trust in persistent-memory agents is widespread while net harm depends on model capability, with rigorous but synthetic-task evidence limited to 0.6-8B models; reproducibility readiness is rated high.
Quantization damage is diffuse — spend the next bit globally
A well-controlled causal study finds no cheap localizer predicts quantization damage and that global g128 granularity outperforms oracle local repair on models up to 8B parameters; reproducibility readiness is rated medium.

Timeline

2023OpenAlex records the earliest activity in the snapshot: 1 publication in 2023, which also marks the first work in the counts-by-year record.
2026Output accelerates to 7 works in 2026, the strongest year on record and the basis of the 61/100 activity score.
2026arXiv preprint on the structure of quantization damage in LLMs: damage is diffuse, no cheap localizer predicts it, and global g128 granularity beats oracle local repair (models up to 8B).
2026arXiv preprint The Memory Trust Gap shows stale-memory over-trust in persistent-memory agents is capability-dependent, with evidence on 0.6-8B models.
2026arXiv preprint XMerge proposes cross-axis selection and reconstructive layer merging for LLM depth compression, evaluated across 7 models.

Papers in the Evolve catalog

Frequent co-authors

Jundong Hu×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.

Bibliometric indicators (h-index, i10-index, per-year counts, topics, distinct coauthors) come from an OpenAlex snapshot only; no other bibliometric source is used.
No institution, country, or ORCID is available; the profile is name-resolved from public bibliographic data and may be affected by namesake ambiguity.
The Evolve catalog covers only three 2026 arXiv papers; the author's other outputs were not independently reviewed.
Output, impact, focus, and reproducibility axes are unscored because underlying data is insufficient; the citation record (h-index 1) is too thin for impact claims.