Geetanjali Bihani

Geetanjali Bihani

PhD Candidate

Purdue University

I am a recent Ph.D. graduate from Purdue University, where I worked on Natural Language Understanding with Dr. Julia Taylor Rayz at the AKRaNLU Lab. My research focuses on enhancing the reasoning capabilities of language models by tackling linguistic ambiguity and improving generalization, especially in cases where the structural form (locution) differs from the intended communicative function (illocution), as often seen in manipulative discourse.

During my Ph.D., I explored a range of challenges in language model reasoning. I developed methods to quantify how alignment methods exacerbate opposing biases, demonstrated that calibrated models lack generalizability, developed methods for sense-enrichment of contextualized representations, worked on fuzzy models for intent classification, including task-based intent as well as covert malicious intents.

My research takes a two-pronged approach:

  • Structured Knowledge for Context-Aware Detection: Develop and refine structured representations of language to help models better understand social context, with a focus on detecting manipulative discourse.
  • Uncertainty Estimation in Language Tasks: Develop methods to quantify the uncertainty in language model outputs when reasoning with underspecified inputs.

For more information regarding my work, please check out my Google Scholar page.

Recent Updates

  • Mar 2025 - Received A.H. Ismail Interdisciplinary Graduate Degree Grant from Purdue University!
  • Mar 2025 - Received Graduate Student Travel Grant from Purdue Polytechnic Institute!
  • Jan 2025 - Delivered an ignite talk, “Bridging the Gap: Advancing AI for Detecting Covert Online Harms,” in the Digital and Social Media Track at HICSS-58
  • Dec 2024 - Paper on “Examining Language Model’s Behavior with Occupation Attributes” accepted to COLING 2025 - paper
  • Aug 2024 - Paper on “The Reliability Paradox: Exploring How Shortcut Learning Undermines Language Model Calibration” accepted to HICSS-58 - paper

Contact