Reshabh K Sharma

I am a PhD student at Paul G. Allen School of Computer Science & Engineering . I am advised by Professor Dan Grossman. I work at the intersection of Programming Languages (PL) and LLMs. My research focuses on the development of infrastructure and tools to improve the reliability, security, and maintainability of LLM-based software, leveraging best practices from traditional software engineering.
Need for infrastructure and tools for LLM-based software: Maintaining LLM-based software is challenging because changes to system prompts or agent configurations can trigger unpredictable regressions that are difficult to detect and reproduce. The non-deterministic outputs of LLMs mean that even minor prompt edits or model updates may silently break functionality, degrade user experience, or introduce security risks. As systems scale and rely on complex prompt chains and multiple agents, tracking dependencies and ensuring consistent behavior becomes increasingly difficult without dedicated infrastructure for regression testing and prompt management.
Usefulness of specifications for improving the reliability of LLM-based software: Specifications are essential in LLM-based software because they define clear expectations for how prompts, agents, and systems should behave, enabling automated validation and enforcement. In SPML, the program acts as a specification to detect deviations at runtime and validate various types of input, including images. PromptPex uses specifications to generate targeted tests, and agent access control systems use them to enforce permissions during execution. By formalizing requirements, specifications help make LLM-driven systems more robust, secure, and maintainable.

Automatic Test Generation for Language Model Prompts

Reshabh K Sharma, Jonathan De Halleux, Shraddha Barke and Benjamin Zorn. PromptPex: Automatic Test Generation for Language Model Prompts arXiv preprint arXiv:2503.05070 (2025).

Blog: Testing AI Software Isn't Like Testing Plain Old Software (Reshabh K Sharma, Peli de Halleux, Shraddha Barke and Ben Zorn)

Blog: Prompts are Programs (Tommy Guy, Peli de Halleux, Reshabh K Sharma, and Ben Zorn)

Prompt Specification Language for protecting LLM-based systems

Reshabh K Sharma, Vinayak Gupta, and Dan Grossman. SPML: A DSL for Defending Language Models Against Prompt Attacks. arXiv preprint arXiv:2402.11755 (2024)

SPML Prompt Injection Dataset: A comprehensive dataset for research on prompt injection attacks and defenses in LLM-based systems, with over 5,000 downloads. Available at Hugging Face and the project page.

Reshabh K Sharma, Vinayak Gupta, and Dan Grossman. Defending Language Models Against Image-Based Prompt Attacks via User-Provided Specifications SAGAI at IEEE S&P (2024).