Submitted the chain-of-thought faithfulness paper to ACL Rolling Review (October 2026 cycle).
Computer Science and Mathematics at Southern Miss
SuramyaAngdembay
I'm an undergraduate researcher at The University of Southern Mississippi working on mechanistic interpretability, chain-of-thought faithfulness, and trustworthy AI. My research asks what language models actually do when we evaluate them, using correctness-stratified audits, sparse autoencoders, and causal interventions.
I work with Dr. Nick Rahimi on chain-of-thought faithfulness (CRA UR2PhD program), Dr. Haiyan Tian on a mechanistic audit of insider-threat language models, and Dr. Gregg Jacobs on error analysis of NOAA's RTOFS ocean forecasts. More about my research.
Recent
News
arXiv v2 of Two Regimes of Chain-of-Thought Unfaithfulness; analysis code and results released on Zenodo.
First preprint on arXiv: arXiv:2607.23458, with D. Aryal and N. Rahimi.
Selected as a Break Through Tech AI Fellow (Cornell Tech).
Talk at the 33rd USA/USM/SELU Research Mini-Conference: continuous-time Markov chains for insider-threat behavior modeling.
Joined the Institute for Advanced Analytics and Society to work with Dr. Gregg Jacobs on NOAA RTOFS ocean forecasts.
ACM SIGHPC Immersion Program at SC25, St. Louis (one of 15 students selected nationally); Student HQ lightning talk.
First conference talk: AMS Spring Southeastern Sectional Meeting, Clemson University.
Preprints and manuscripts
Papers
Two Regimes of Chain-of-Thought Unfaithfulness: Metric-Based Detection Fails Where Models Are Wrong
S. R. Angdembay, D. Aryal, and N. Rahimi. arXiv preprint; under review at ACL Rolling Review.
arXiv:2607.23458 Code and dataWhat Do Benign-Trained Session Language Models Actually Detect? A Mechanistic Audit on Insider-Threat Benchmarks
S. R. Angdembay and H. Tian. Manuscript in preparation, 2026.