← All projects
LLM · Research

Code Reasoning & Test Generation

Prompt engineering for code-specialized LLMs

An exploration of how prompting strategy shapes a code LLM's ability to reason about programs and generate useful unit tests, using DeepSeek-Coder as the base model.

Type
ML / research
Model
DeepSeek-Coder
Focus
Prompt engineering
Language
Python
Code reasoning & test generation

01 Overview

Code LLMs are sensitive to how a task is framed. This project designed and compared tailored prompt-engineering strategies aimed at two related goals: improving the model's reasoning about a given piece of code, and improving the quality of the unit tests it generates for that code.

02 What I worked on

  • Built and iterated on prompt templates for code reasoning and test synthesis.
  • Generated tests against target functions and evaluated them for correctness and coverage.
  • Compared strategies to see which framings produced more reliable model behavior.

More detail and results available on request.