Early-career researcher in efficient LLM inference, small language model pretraining, and evaluation reliability.
3
papers here
2
reviewed by Evolve
3
first-author
1
corresponding
—
works · OpenAlex
0
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—
Output cannot be scored yet: the profile records only 4 works in the last three years and no complete publication list. With a fuller indexed corpus, this axis would measure annual productivity and cumulative publication volume.
Impact—
Impact data is insufficient: OpenAlex reports h-index 0 and i10-index 0, consistent with very recent arXiv-only output rather than measured influence. This axis would measure citation-based reach — total citations, citing works, and h-index growth once citations accrue.
Focus0/100
The focus score is 0, reflecting an even 0.5/0.5 split across two field labels — machine learning systems (efficient LLM inference) and natural language processing (small language model pretraining) — and five OpenAlex topic assignments led by Topic Modeling (4) and Natural Language Processing Techniques (4). The three reviewed papers nonetheless share one practical thread: efficient, trustworthy LLM serving and evaluation.
Collaboration22/100
Collaboration scores 22, matching a small footprint: only 2 distinct coauthors are recorded, with Nirdesh Chauhan and Siddharth Kosaraju as the recurring partners. All three cataloged papers are first-authored, and one carries corresponding authorship, suggesting a compact, hands-on team rather than a large distributed consortium.
Activity47/100
Activity scores 47 on a sparse but current record: OpenAlex counts 1 work in 2023 and 4 in 2026, with recent3yWorks at 4. The 2026 burst comprises three first-authored arXiv preprints spanning KV-cache serving, sub-150M pretraining, and verifier reliability — an active recent cadence with a visible gap before it.
Reproducibility100/100
Reproducibility scores a perfect 100, the strongest signal in this profile: the mean repro-readiness score (reproMean) is 100 across 2 repro-scored papers, and all three Evolve-reviewed papers carry a high repro-readiness rating. Evolve's verdicts repeatedly note candid, honest reporting — ablation ladders, disclosed single-seed limits, and explicitly flagged thin evidence.
Research summary
Dushyant Rajput is an early-stage researcher working at the intersection of efficient LLM inference systems and small language model pretraining, with a growing interest in evaluation reliability. The profile — name-resolved from OpenAlex bibliographic data — records an h-index and i10-index of 0, 4 works in the last three years, and a per-year distribution of 1 work in 2023 and 4 in 2026. Field distribution is split evenly (0.5/0.5) between machine learning systems (efficient LLM inference) and natural language processing (small language model pretraining). All three papers Evolve has reviewed are 2026 arXiv preprints with Rajput as first author, one also as corresponding author. The first quantifies quality and serving tradeoffs of shared-prefix KV-cache reuse across standard LoRA adapters on Qwen3-1.7B; the second runs a controlled 53M-to-110M scaling study of the deep-thin JugnuLM models with a Muon optimizer, value residuals, and distillation; the third measures how cost-saving verifier cascades hide rising error, reported as 3% versus 32%. Recurring collaborators are Nirdesh Chauhan and Siddharth Kosaraju among just 2 distinct coauthors recorded. The standout metric is reproducibility: a perfect 100 mean across 2 repro-scored papers, with all reviewed work rated high for repro readiness. Output and impact dimensions currently lack sufficient citation data and are left unscored.
Collaboration pattern
The recorded collaboration footprint is compact: 2 distinct coauthors, with Nirdesh Chauhan and Siddharth Kosaraju as recurring partners. Rajput takes first-author position on all three Evolve-reviewed papers and corresponding authorship on the KV-reuse study, consistent with a small hands-on team driving each project end to end. The collaboration dimension scores 22, reflecting a narrow but concentrated network; no large consortium or cross-institution pattern appears in the available data.
Research keywords
Highlights
KV-cache reuse across standard LoRA adapters
Controlled scaling below 150M parameters
Reliability cost of cheap verifier cascades
Timeline
- 2023Earliest recorded output in the OpenAlex yearly counts: 1 work.
- 2026First-authored arXiv study on shared-prefix KV reuse across standard LoRA adapters (also corresponding author), rated high for reproducibility.
- 2026First-authored scaling study of JugnuLM-53M versus JugnuLM-110M in the sub-150M regime.
- 2026First-authored measurement paper on the reliability cost of cost-saving verifier cascades, closing a 4-work year in the OpenAlex counts.
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.
- Bibliometrics come from OpenAlex only (identity: openalex); no institution, country, or ORCID is verified, and the profile is name-resolved from public bibliographic data.
- h-index (0) and i10-index (0) reflect very recent arXiv-only output; citation-based impact cannot yet be assessed.
- OpenAlex topic labels include off-topic entries (e.g., Craniofacial Disorders and Treatments), which may indicate name-resolution noise for a shared name.
- Qualitative verdicts rest on the 3 Evolve-reviewed papers; the reproducibility score (reproMean 100) is computed from 2 repro-scored papers.