Publications
💡 Hover over nodes to highlight connections. Click to filter the papers list.
VATS: Exploiting Implicit Authority in Error-Path Injection via Systematic Mutation
Spotlight
workshop
Harshil Patel, Kunal Pai
Second Workshop on Agents in the Wild: Safety, Security, and Beyond, ICML 2026
This paper introduces VATS, a framework demonstrating that autonomous agents are highly vulnerable to "error-path injections," where adversarial payloads disguised as tool error messages exploit the Model Context Protocol (MCP) to bypass safety heuristics and triple the success rate of standard indirect prompt injections across major frontier models.
Artificial IntelligenceLarge Language Models (LLMs)AI Safety
Toward Reproducible and Standardized Computer Architecture Simulation with gem5
conference
Kunal Pai, Harshil Patel, Erin Le, Noah Krim, Mahyar Samani, Bobby R. Bruce, Jason Lowe-Power
IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) 2026
To address the inconsistencies in simulation-based research, this work enhances the gem5 ecosystem by standardizing disk-image creation, introducing a flexible class-based exit event system for better guest-host communication, and implementing native tools like Suites and MultiSim to streamline and stabilize complex multi-workload workflows.
Computer Architecturegem5SimulationReproducibility
NAAMSE: Framework for Evolutionary Security Evaluation of Agents
workshop
Kunal Pai, Parth Shah, Harshil Patel
ICLR 2026 Agents in the Wild: Safety, Security, and Beyond Workshop
NAAMSE is an evolutionary framework that automates AI agent security testing by using a feedback-driven optimization process to mutate prompts and uncover high-severity vulnerabilities that manual or static benchmarks often miss.
Artificial IntelligenceLarge Language Models (LLMs)AI SafetyAdversarial Testing
Implications of Full-System Modeling for Superconducting Architectures
workshop
Kunal Pai, Mahyar Samani, Anusheel Nand, Jason Lowe-Power
Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC Workshops '25)
As Moore's Law slows, superconducting electronics offer ultra-low-power, high-speed computation potential. This paper presents the first full-system superconducting modeling in gem5, including cryogenic and superconducting cores, caches, and interconnects. Our results show that superconducting cores and caches can yield up to 24× speedup for compute-intensive workloads, but memory-intensive applications remain bottlenecked by room-temperature DRAM. This makes superconducting technology more suitable for domain-specific accelerators rather than general-purpose computing, with performance dependent on workload memory access patterns and data widths.
Computer ArchitectureSuperconductingCryogenic Computinggem5
HASHIRU: Hierarchical Agent System for Hybrid Intelligent Resource Utilization
preprint
Kunal Pai, Parth Shah, Harshil Patel
arXiv preprint
To support resource-efficient multi-agent reasoning, we introduce HASHIRU, a hierarchical agent system that dynamically instantiates specialized agents under cost and memory constraints. HASHIRU combines hybrid LLM usage, autonomous API/tool creation, and a novel economic model for agent hiring/firing, outperforming larger models like Gemini 2.0 Flash on complex reasoning and safety tasks.
Artificial IntelligenceLarge Language Models (LLMs)Multi-Agent Systems
CoDocBench: A Dataset for Code-Documentation Alignment in Software Maintenance
conference
Kunal Pai, Premkumar Devanbu, Toufique Ahmed
International Conference on Mining Software Repositories (MSR) 2025: Data and Tool Showcase Track
Understanding and implementing code changes is a key aspect of software maintenance. To support this, we introduce a new dataset of coupled changes to code and documentation mined from high-quality GitHub projects, where each sample represents a single commit with simultaneous updates to code and docstrings. This dataset enables training and evaluation on realistic, change-related tasks, which remain challenging for current models like Llama 3.1 405B and Mixtral 8×22B.
Software EngineeringGitHub MiningLarge Language Models (LLMs)
Calibration and Correctness of Language Models for Code
conference
Claudio Spiess, David Gros, Kunal Suresh Pai, Michael Pradel, Md Rafiqul Islam Rabin, Amin Alipour, Sushmit Jha, Premkumar Devanbu, Toufique Ahmed
International Conference on Software Engineering (ICSE) 2025
Machine learning models often produce incorrect outputs, making reliable confidence measures essential for determining the trustworthiness of these outputs. This paper introduces a framework to evaluate and improve the calibration of code-generating models, finding that these models are generally poorly calibrated initially but can be improved using methods like Platt scaling, thereby enhancing decision-making in software engineering.
Software EngineeringMachine LearningNaturalness of Software
Potential and Limitation of High-Frequency Cores and Caches
poster
Kunal Pai, Anusheel Nand, Jason Lowe-Power
ModSim 2024: Workshop on Modeling & Simulation of Systems and Applications
The poster presentation explores the potential and limitations of high-frequency in-order and out-of-order cores and caches in modern processors, highlighting the trade-offs between speedups and bandwidth.
Computer ArchitectureCryogenic ComputingSuperconducting
Automatic semantic augmentation of language model prompts (for code summarization)
conference
Toufique Ahmed, Kunal Suresh Pai, Premkumar Devanbu, Earl T. Barr
International Conference on Software Engineering (ICSE) 2024
Adding explicit semantic facts as prompts to Large Language Models improves their performance in code summarization tasks, with notable improvements exceeding 2 BLEU and, in some cases, even surpassing 30 BLEU, demonstrating the effectiveness of this approach in enhancing code analysis and extraction of essential information.
Software EngineeringMachine LearningNaturalness of Software
Validating Hardware and SimPoints with gem5: A RISC-V Board Case Study
poster
Kunal Pai, Zhantong Qiu, Jason Lowe-Power
gem5 Workshop at International Symposium on Computer Architecture (ISCA) 2023
The poster discusses the development of a RISC-V board model (RISCVMatched) in gem5, along with a methodology for fine-tuning gem5 configurations to closely match real-life systems, resulting in more accurate hardware validation and simulation capabilities.
Computer Architecturegem5
gem5 Vision
poster
Parth Shah, Kunal Pai, Harshil Patel, Arslan Ali
gem5 Workshop at International Symposium on Computer Architecture (ISCA) 2023
The gem5 Vision Project seeks to improve user-friendliness and accessibility by introducing advanced search functionality, comprehensive resource categorization, and expanded database support within the gem5 ecosystem for researchers and developers.
Computer Architecturegem5
✨ Citation copied to clipboard!