Arush Sharma

Research engineer · IIT (ISM) Dhanbad ’27

Arush Sharma

I work on making large models faster and more reliable: draft-and-verify decoding for diffusion LLMs, structure-aware KV caches, post-training, and uncertainty-aware representation learning.

Final-year Electronics & Communication student at IIT (ISM) Dhanbad. Right now I’m a Research Engineer Intern at Conscious Engines. Before that I was in the AI-for-Quantum group at IBM Research, where my Sign-hedged Beam QAOA work won the Best Paper Award at the 2026 internship symposium.

Email me GitHub LinkedIn Hugging Face
Best Paper IBM Research Internship Symposium 2026
5 / 80,000 teams, AMLC 2026 entity resolution
2 papers under review at WACV & AAAI 2027
AIR 5181 JEE Advanced, top 0.47% of 1.11M+

Experience

  1. Conscious Engines Stackr Labs India · backed by a16z

    Aug 2026 – Present

    Research Engineer Intern

    • Built draft-aligned decoding for SDAR-30B-A3B, a diffusion mixture-of-experts LM: a dense 1.7B SDAR drafter pre-fills each 32-token block, and a self-distilled 26M-parameter LoRA judge commits or re-masks every token.
    • Added a 35M LoRA (attention + MLP) that aligns the drafter with the 30B MoE, cutting passes per block from 12.8 to 7.6 for a 1.35× throughput gain at comparable accuracy across benchmarks.
    • Scaling FLARE-2B (Qwen3.5-2B, autoregressive-to-diffusion): token-balanced training through 9.66B tokens on 8×H200, then 7×RTX PRO 6000 Blackwell at full utilisation, with resumable DCP checkpoints and W&B.

  2. IBM Research

    May 2026 – Aug 2026

    Research Intern, AI for Quantum · with Dr. Anupama Ray

    • Built a CliffSearch-style multi-agent algorithm-discovery loop where open- and closed-model agents propose, mutate, critique and evaluate analytic QAOA schedule families over many generations.
    • Evolved the linear-ramp baseline into a fixed power-law schedule, carried it from plain QAOA into LR-QAOA/RQAOA, and paired it with sign-hedged beam recursion across simulator, MPS and IBM quantum hardware.
    • The resulting Sign-hedged Beam QAOA won the Best Paper Award at the IBM Research Internship Symposium 2026.

  3. IIT (ISM) Dhanbad

    Jun 2025 – Present

    Research Intern · with Dr. Soumen Basu & Dr. Manoj Kumar

    • Self-supervised learning for medical imaging and graph models for brain networks; two papers under review, described in Research.

  4. Aeroqube Inc.

    Jun 2025 – Nov 2025

    Agentic AI Intern

    • Fine-tuned MedGemma 4B for a medical-tourism assistant and deployed it as a Dockerised AWS pipeline with RAG, LangSmith tracing and evaluations, and repeatable prompt/model comparisons.
    • Generated personalised health itineraries with the Gemini API, grounded through LangGraph RAG on MedGemma context, behind a Vercel-hosted frontend.

Research

Under review · WACV 2027 First author

From Points to Distributions: Probabilistic Latent Autoregression for Medical Image Representation Learning

Deterministic self-supervised learning collapses ambiguous anatomy into over-confident point embeddings. PACL predicts a Gaussian latent per patch instead and carries that uncertainty into classification and segmentation. I formulated the mean/variance and anti-collapse losses.

  • 0.987 CXR AUROC
  • 0.938 BUSI AUROC
  • 0.604 IoU
  • +2% / +2.1% / +1.3 pts over prior SOTA

With Dr. Soumen Basu (Meta) and Dr. Manoj Kumar (IIT Roorkee)

Under review · AAAI 2027 Second author

BrainRMA: ROI-Preserving Multiscale Attention for Brain Network Analysis

Brain-graph models either miss long-range dependencies or erase region identity through pooling. BrainRMA keeps ROI tokens and fuses global attention with reliability-weighted ROI-to-cluster context. I ran the ablation suite and co-wrote the methodology and experiments.

  • 82.35% ABIDE
  • 93.75% OSF
  • +3.35 / +6.25 pts over SOTA

With Dr. Soumen Basu (Meta) and Dr. Manoj Kumar (IIT Roorkee)

In preparation Best Paper · IBM Research

Sign-hedged Beam QAOA: Scaling Quantum Optimization Algorithms

Pairs a CliffSearch-evolved power-law schedule with width-B sign hedging, so recursive QAOA no longer makes irreversible correlator-sign errors.

  • 0.998 SK ±1 (n=20), 2.1× AR
  • 1.000 on IBM Marrakesh
  • 0.984 MaxCut n=60, +17.6% vs QAOA
  • 0.983 20-asset portfolio, +26% vs RQAOA

Projects

PriorityKV

GitHub ↗

Structure-aware KV-cache eviction for long agent traces

Protects system, tool-schema, constraint, sink and recent tokens, falling back to attention scores only when protected context exceeds the budget. Over 15,360 generations on Qwen3-8B and Llama-3.1-8B it kept task state in 112/120 cases versus 1/120 for position-blind eviction.

−28% KV payload · −13% peak CUDA memory · 11–12% latency cost, measured on H200 in BF16/INT4

  • PyTorch
  • Transformers
  • H200
  • INT4

Business Entity Resolution

GitHub ↗

AMLC 2026 · rank 5 of 80,000 teams

A cost-tiered matching cascade over 1.7M businesses in the US, India and France. A fine-tuned multilingual-e5 bi-encoder blocks to 20 candidates at 99.5% recall; four cross-encoders and LightGBM resolve 93% of cases; six Qwen3.5-4B LoRA judges handle the uncertain 6%.

0.99151 public leaderboard · held-out F0.5 0.9906 → 0.9928 · 5.09B params under the 8B limit

  • multilingual-e5
  • bge-reranker
  • LoRA
  • LightGBM

ReVival

Demo ↗ GitHub ↗

Amazon HackOn Season 6 · circular-commerce agents

Routes returned products back to market instead of landfill with a 7-agent pipeline: Bedrock vision grading, disposition, pricing, buyer matching and S3-hosted “Trust Passport” condition reports. I owned the supply-side FastAPI backend, 7 DynamoDB tables and a Titan embedding cache that skips redundant grading calls.

  • FastAPI
  • AWS Bedrock
  • DynamoDB
  • Next.js

Arista Intelligent RRM

GitHub ↗

Inter IIT Tech Meet 14.0 · 12th of 23 IITs

A three-loop radio-resource manager that turns SDR measurements into interference graphs, then uses a GNN state encoder with offline Conservative Q-Learning to tune channel, power, bandwidth and OBSS-PD on Linux virtual APs.

~18% edge-client throughput · zero configuration churn · P95 retries under 8%

  • GNN
  • Offline RL
  • SDR

Hindi TinyStories GPT

GitHub ↗

A 33M-parameter GPT written from scratch

Decoder-only transformer with RoPE, RMSNorm and SwiGLU, every component hand-written in PyTorch and trained on ~100M tokens of Hindi children’s stories. It starts its stories with “एक बार की बात है…”

val perplexity 2.26 · 197M tokens seen · 98 minutes on a single L4

  • PyTorch
  • From scratch
  • bf16

AtomMail

GitHub ↗

Hackfest · 2nd prize

An email assistant that reads threads and attachments, OCRs screenshots and PDFs, retrieves the user’s history from MongoDB and drafts replies in their tone with LangChain over Gemini and Hugging Face models. Cut end-to-end reply time by about 80%.

  • Flask
  • LangChain
  • Tesseract
  • MongoDB

More from GitHub

  • AlignmentLabMatched-compute comparison of SFT, DPO and GRPO/RLVR on Qwen3-8B and 14B: does RLVR add capability, or only sharpen sampling?
  • DeepLense (GSoC)Physics-informed gravitational-lens classification and finding; 0.9976 ROC-AUC, weights public on Hugging Face.
  • Mellea workspaceStructured, testable LLM programs with IBM’s Mellea: ReAct agents, validators and MCP demos.
  • DPO on a budgetPreference optimisation with DPO + QLoRA under tight hardware limits.
  • Stanford CS336Tokenizers, transformers and training loops, implemented from the ground up.
  • Mistral summarisationLoRA/QLoRA fine-tuning with ROUGE and BERTScore evaluation.

Toolbox

Languages
Python, C/C++, SQL, JavaScript/TypeScript
Training
PyTorch, Transformers, TRL/PEFT, OpenRLHF, vLLM, PyTorch Geometric, scikit-learn
Methods
SFT, DPO, GRPO/RLVR, LoRA/QLoRA, KV-cache optimisation, quantisation, ablations, OOD evaluation
Systems
FlashAttention/Triton, CUDA, DDP/FSDP, LSF, Docker, AWS (Bedrock, S3, DynamoDB)
Apps
FastAPI, Flask, Next.js/React, LangChain, LangGraph, MongoDB

Honours

  • Best Research Paper AwardIBM Research India · Internship Symposium 2026
  • Rank 5 of 80,000 teamsAMLC 2026 · Business Entity Resolution
  • 2nd PrizeHackfest · AtomMail
  • 12th of 23 IITsInter IIT Tech Meet 14.0 · Arista RRM
  • AIR 5181JEE Advanced · top 0.47% of 1.11M+
  • Amazon ML Summer SchoolSelected, top 3,000 of 134,000
  • DSA CertificateGoogle Student Upskill Launchpad

Education

IIT (ISM) Dhanbad

B.Tech, Electronics and Communication Engineering · CGPA 8.19 / 10

Expected April 2027

Let’s talk.

I’m interested in research engineering roles and collaborations on efficient inference, post-training and evaluation. The fastest way to reach me is email.

arushsharma536@gmail.com