Efficient-inference researcher ternarising Qwen3 LLMs — 2026 arXiv work on 1.58-bit post-training quantization
3
papers here
2
reviewed by Evolve
0
first-author
0
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—
No output score is available: the snapshot provides no validated publication or venue counts for this name, so this axis — which would measure total and per-year scholarly output — is left unscored pending more complete records.
Impact—
Impact cannot be scored yet: h-index and i10-index are recorded as 0 with no citation time series, so this axis — which would measure citations, influence, and uptake of the ternary-quantization work — is not yet meaningful.
Focus0/100
The focus axis scores 0/100: OpenAlex distributes this record across four topic clusters — Topic Modeling and Advanced Neural Network Applications (3 works each), Low-power high-performance VLSI design (2), and Domain Adaptation and Few-Shot Learning (1) — while the field split is an even 0.5/0.5 between two near-identical efficient-inference labels. The catalog itself is tightly concentrated on ternary quantization, so the dispersion partly reflects classifier noise around a young profile.
Collaboration22/100
Scoring 22/100, collaboration is the profile's narrowest measured axis: the OpenAlex snapshot lists only 2 distinct co-authors, with Anirudh Malik and Poojith Devan as recurring partners. All three catalogued papers carry middle authorship without corresponding-author status, indicating contribution within a small, stable team rather than an independent or large-scale collaborative network. Expanding the co-author base and leading authorship would move this axis most.
Activity40/100
Activity scores 40/100: all 3 works in the last three years fall in 2026 (countsByYear: 2026 — 3), giving a compact but very recent burst rather than a sustained multi-year record. The three catalogued arXiv papers — Qwen3-4B ternarization, Qwen3-8B scaling, and capability-stratified degradation — were all posted in 2026. Whether this becomes steady output is the key open question for the profile.
Reproducibility55/100
Reproducibility scores 55/100, averaged over 2 papers (reproMean 55, reproPapers 2), both rated medium readiness. Evolve's review credits the Qwen3-8B work with a rare reproduction gate and candid reporting, and the Qwen3-4B study with transparent accounting of 1.641 bits/weight, a 9.8-point accuracy drop, and 8.29→3.96 GiB storage. Reported limits — single-seed n=500 evaluations, no competing baselines, a 7.9% FP16 gap — cap the score.
Research summary
M Sparsh Mehra is an early-stage machine-learning researcher whose visible record is concentrated in one niche: post-training ternarisation of large language models. Three arXiv papers, all dated 2026, form a coherent sequence — a Qwen3-4B ternarization study reporting 1.641 bits per weight, a 9.8-point accuracy drop, and storage shrinking from 8.29 to 3.96 GiB; a scale-up to Qwen3-8B covering capability retention, reproduction, lossless packing, and packed execution; and an analysis of capability-stratified degradation in ternary language models. Evolve's review of the 8B paper highlights a rare reproduction gate and candid reporting, while flagging single-seed n=500 evaluations, absent competing baselines, and a 7.9% FP16 gap; the 4B study is judged transparent but baseline-light. Bibliometrics, drawn from an OpenAlex snapshot, are those of a brand-new record: h-index 0, i10-index 0, 3 works in the last three years, and 2 distinct co-authors — Anirudh Malik and Poojith Devan — with middle authorship and no corresponding role on catalogued papers. OpenAlex topics disperse across topic modeling, advanced neural network applications, low-power VLSI design, and few-shot learning, even though the field split sits entirely within efficient inference and model compression. The profile is name-resolved from public bibliographic data; no affiliation or ORCID is linked.
Collaboration pattern
The OpenAlex snapshot records only 2 distinct co-authors, with Anirudh Malik and Poojith Devan recurring across the catalog. All three papers list the researcher in a middle authorship position without corresponding status, consistent with a small, stable three-person team producing a focused paper series rather than a broad collaborative network. The collaboration score of 22/100 reflects this narrow footprint.
Research keywords
Highlights
Scaling ternarisation to Qwen3-8B with a reproduction gate
Qwen3-4B ternarized to 1.641 bits per weight
Capability-stratified degradation in ternary LMs
A third 2026 arXiv paper analyzes how degradation varies across capability strata in ternary language models; it is catalogued but not yet reviewed by Evolve, so no verdict or reproducibility score exists.
Timeline
- 2026Posted the Qwen3-4B post-training ternarization study (1.641 bits/weight; storage 8.29→3.96 GiB).
- 2026Followed with the scale-up to Qwen3-8B: capability retention, reproduction, lossless packing, and packed execution.
- 2026Published an analysis of capability-stratified degradation in ternary language models.
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 and scores derive from an OpenAlex snapshot only; no ORCID or institutional affiliation is linked, so identity is name-resolved from public bibliographic data.
- Citation metrics (h-index 0, i10-index 0) reflect a very young 2026 record; all works are recent and citations have not yet accrued.
- The third catalogued paper, 'Capability-Stratified Degradation in Ternary Language Models', has no Evolve verdict, keywords, field label, or reproducibility scoring.
- OpenAlex topic and field labels are automatic classifications and may not match self-described research areas.