Publications

Incoming CS PhD student at Dartmouth College (fall 2026), advised by Prof. Shawn Shan. Research spans ML security, LLM privacy, low-resource NLP, and production AI systems. Also on Google Scholar.

Published & Accepted

SLMs as Compiler Experts: Auto-Parallelization for Heterogeneous Systems

Published

NeurIPS 2025 · Workshop on Machine Learning for Systems

Prathamesh Devadiga, et al.

Small Language Models as compiler experts for auto-parallelization in heterogeneous systems. Achieves 6.81× average speedup (43.25× peak) against LLVM Polly, TVM, and Triton baselines.

Compiler Optimization SLMs Auto-Parallelization Heterogeneous Systems

Can LLMs Learn Tulu? Teaching Low-Resource Languages Through Hard Constraint Prompting

Published

EACL 2025 · LoResLM Workshop

Prathamesh Devadiga, et al. · Lossfunk

Hard negative constraints reduce catastrophic language leakage from 80% to 5% in low-resource language modeling for Tulu (~0.001% of typical training data). Explicit prohibitions outperform positive instructions for maintaining language integrity.

Low-Resource NLP Constraint Learning Indic Languages

GUARDIAN: Multi-Agent Defense System for Large Language Model Security

Published

ICIAI · Waseda University

Prathamesh Devadiga, et al.

Multi-agent defense system achieving 99.5% prompt injection and jailbreak detection accuracy through coordinated multi-agent analysis.

LLM Security Multi-Agent Systems Jailbreak Detection

PyraFuseNet: Dual-Path Network for Resource-Constrained Vision

Published

ICIAI · NTU Singapore

Prathamesh Devadiga, et al.

Dual-path architecture for resource-constrained vision. Achieves 55% computational reduction vs. ResNet-18 while maintaining competitive performance.

Computer Vision Efficient Networks Neural Architecture

PDF Malware Detection with Adversarial Robustness

Published

Applied Soft Computing · Q1 Journal

Prathamesh Devadiga, et al. · IIT Indore

KASPER framework for PDF malware detection: 99.5% accuracy under FGSM and PGD adversarial attacks. Explainability via Kolmogorov-Arnold Networks.

Malware Detection Adversarial Robustness Deep Learning

Under Review

Breaking the Query Barrier: Efficient Verbatim Regurgitation from Large Language Models

Under Review

ICML 2026 · Main Conference

Prathamesh Devadiga, Shawn Shan, A. Krishna · Dartmouth College

Hierarchical Extraction Search (HES): a query-efficient framework extracting memorized training data from LLMs under black-box constraints. Reduces extraction cost by 10–100× vs. brute-force, demonstrated across 12 models (7B–123B). Models appearing robust to high-volume attacks remain vulnerable to low-footprint extraction.

LLM Privacy Memorization Data Extraction ML Security

Preprints

SAMVAD: Multi-Agent Framework for Modeling Judicial Reasoning in Indian Jurisprudence

Preprint

arXiv · 2509.03793

Prathamesh Devadiga, et al.

Multi-agent framework for modeling judicial reasoning in Indian jurisprudence.

Legal AI Multi-Agent Judicial Reasoning

arXiv ↗

RegimeNAS: Regime-Aware Neural Architecture Search for Financial Trading

Preprint

arXiv · 2508.11338

Prathamesh Devadiga, et al.

Regime-aware NAS for financial trading, adapting architecture selection to market conditions.

Neural Architecture Search Financial ML Regime Detection

arXiv ↗

MorphNAS: Differentiable Neural Architecture Search for Multilingual NER

Preprint

arXiv · 2508.15836

Prathamesh Devadiga, et al.

Differentiable NAS for multilingual named entity recognition across multiple languages.

Neural Architecture Search Multilingual NLP NER

arXiv ↗