research interests

I'm interested in exploring methodologies that help us understand and bridge the socio-technical gap in AI systems—particularly in how they are evaluated, deployed, and governed.

Currently thinking about:

 

• Improving Evaluations e.g. How can we better measure evaluations and datasets to ensure construct and claim validity?
• Context-specificity e.g. How do we operationalize evaluations in multilingual and code-switching environments?
• Performance Robustness e.g. To what extent are model behaviors stable—and hence evaluation results robust—across perturbations, shifts, or rephrasings?

 

If you're also thinking about the above, let's chat.

selected projects & work

• CALMA: Context-Aligned Language Model Alignment

  Prajna Soni, Deepika Raman, Dylan Hadfield-Menell — Preprint @ ICML 2025 Technical AI Governance Workshop ** Oral Presentation **

  A framework for deriving context-specific values and alignment axes for language models by leveraging social research theories.

• Addressing Misalignment in Language Model Deployments through Context-Specific Evaluations

  Prajna Soni — SM Thesis @ MIT

  Investigates technical and regulatory methods to assess and mitigate deployment-time misalignment of LLMs.

• FARE: Fair Allocation RE-weighting

  April Chen, Prajna Soni — Paper @ AAAI 2023 AI for Social Good Workshop

  Introduces a re-weighting approach to improve fairness in allocation settings, ensuring equitable outcomes across demographic groups.

• Natural Language Processing: Understanding the Current Landscape

  Medha Patki, Prajna Soni — Article @ MIT Science Policy Review V3

  A policy-oriented overview of contemporary trends and challenges in NLP, including ethical and governance considerations.