Poojith Devan

—

Unclaimed · AI-structured profileOpenAlex

Early-career researcher in ternary post-training quantization of Qwen3 LLMs for efficient inference

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—

Output metrics are not yet scorable: this record shows an h-index of 0, an i10-index of 0, and 4 works counted in 2026 with none earlier. The axis would measure cumulative publication volume and productivity once the record has time to accumulate.

Impact—

Impact cannot be scored yet: with an h-index of 0 and an i10-index of 0, and all works dated 2026, citation signals are at the floor and carry no discriminating information. This axis would measure citation uptake and scholarly influence as the literature accrues.

Focus0/100

The focus axis scores 0, the lowest possible, indicating no single dominant cluster in this record. Field distribution splits evenly between two near-identical labels — machine learning (model compression / efficient inference) at 0.5 and machine learning — efficient inference and model compression at 0.5 — while topic counts spread across five clusters: Advanced Neural Network Applications (4), Topic Modeling (3), Low-power high-performance VLSI design (2), and two singleton topics.

Collaboration22/100

Scoring 22, the collaboration axis reflects a very small working group: only 2 distinct co-authors appear in the record, with Anirudh Malik and M Sparsh Mehra recurring across the papers. All three Evolve-catalogued 2026 papers place this researcher in the last author position and none as corresponding author, consistent with a compact, hands-on team rather than a large distributed consortium.

Activity47/100

With a mid-range score of 47, activity is concentrated entirely in 2026: the year-level counts record 4 works in 2026 and none before, and recent-3-year output stands at 4 works. This pattern suggests an emerging or recently activated publishing rhythm rather than a sustained multi-year cadence, so the score may understate current momentum.

Reproducibility55/100

Reproducibility scores 55 (reproMean 55) across 2 evaluated papers, both rated medium readiness. Evolve's reviews note a rare reproduction gate and candid reporting — transparent disclosure of a 7.9% FP16 gap and of single-seed n=500 evaluations — while flagging the absence of competing baselines. Honesty is a clear strength; external validation remains incomplete.

Research summary

Poojith Devan is an early-stage researcher working on post-training quantization of large language models, with a public record concentrated entirely in 2026. The OpenAlex-derived record lists 4 works in 2026 (recent-3-year works: 4), an h-index of 0, and an i10-index of 0 — floors that reflect a record just beginning to accumulate rather than established impact. Three papers are catalogued by Evolve, all in the last author position and none as corresponding author, co-authored with Anirudh Malik and M Sparsh Mehra (2 distinct co-authors in total). The work centres on ternary weight-only quantization of the Qwen3 family using GPTQ: one study ternarizes Qwen3-4B to 1.641 bits per weight, with a 9.8-point accuracy drop and storage falling from 8.29 GiB to 3.96 GiB; a follow-up scales the approach to Qwen3-8B, covering capability retention, lossless packing, and packed execution while candidly reporting a 7.9% FP16 gap, single-seed n=500 evaluations, and no competing baselines; a third analyzes capability-stratified degradation in ternary language models. Reproducibility scores 55 (reproMean 55) across 2 rated papers, both at medium readiness, with transparency and a reproduction gate flagged as strengths. Topic counts cluster on Advanced Neural Network Applications (4) and Topic Modeling (3), with smaller clusters in low-power high-performance VLSI design (2) and two singletons. This is a focused, transparent, still-thin early profile in efficient LLM inference.

Collaboration pattern

Collaboration is tight and consistent: only 2 distinct co-authors appear in the record — Anirudh Malik and M Sparsh Mehra — and all three Evolve-catalogued 2026 papers place this researcher in the last author position, none as corresponding author. The pattern suggests a small, stable team producing a focused series of quantization studies rather than wide multi-group networking; the collaboration dimension scores 22 on this basis.

Research keywords

post-training quantizationternary weights1.58-bit LLMsGPTQweight-only quantizationmodel compressionefficient inferencecapability retentionlow-bit deploymentQwen3 models

Highlights

Scaling ternary PTQ to Qwen3-8B
Extends post-training ternarisation to Qwen3-8B with capability retention, reproduction, lossless packing, and packed execution. Evolve calls the scale-up rare for its reproduction gate and candor, but flags single-seed n=500 evaluations, no competing baselines, and a 7.9% FP16 gap; reproducibility readiness is medium.
Ternarizing Qwen3-4B to 1.641 bits per weight
A transparent post-training study ternarizes Qwen3-4B to 1.641 bits per weight, with a 9.8-point accuracy drop and storage compressed from 8.29 GiB to 3.96 GiB. Evolve rates the reporting honest but notes the lack of external baselines; reproducibility readiness is medium.
Capability-stratified degradation in ternary LMs
A 2026 paper analyzing degradation in ternary language models stratified by capability. Evolve has not yet published a verdict, keywords, or reproducibility rating for it, so this entry rests on the paper's title alone.

Timeline

2026Published a post-training ternarization study of Qwen3-4B, reporting 1.641 bits per weight, a 9.8-point accuracy drop, and storage reduced from 8.29 GiB to 3.96 GiB.
2026Released Capability-Stratified Degradation in Ternary Language Models, examining how ternary compression affects different capability strata.
2026Scaled post-training ternarisation to Qwen3-8B, addressing capability retention, lossless packing, and packed execution.
2026Bibliometric snapshot of the year: 4 works recorded in 2026, with an h-index of 0, an i10-index of 0, and 2 distinct co-authors.

Papers in the Evolve catalog

Frequent co-authors

Anirudh Malik×3 M Sparsh Mehra×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, topic counts, co-author counts) come from OpenAlex only; no institutional or ORCID identity data was available.
The record is concentrated in a single year: year-level counts show 4 works in 2026 and none earlier, so long-term productivity and citation trends cannot be assessed.
One reviewed paper, Capability-Stratified Degradation in Ternary Language Models, has no Evolve verdict, keywords, or reproducibility rating; its characterization is based on the title alone.
The name Poojith Devan is resolved from public bibliographic data; institution and country are unknown, so disambiguation from same-name authors is not guaranteed.