Publications

My research focuses on improving reliability in LM-based code generation, moving beyond verifying correctness with oracles after code is generated toward early detection during the generation process, with the goal of building reliable and affordable code-generation agents.

• NeurIPS 2026 CCF-A Under Review

  Using Semantic Distance to Estimate Uncertainty in LLM-Based Code Generation

  Weilin He, Arindam Sharma, Cristina David

  Large Language Models show strong performance in code generation, but their outputs lack inherent correctness guarantees. We propose a semantic distance-based uncertainty estimation framework that clusters programs by execution behaviour and models inter-cluster distances to provide a reliable confidence score.

  arXiv PDF Code