2025

1. Preprint

   

   Right Prediction, Wrong Reasoning: Uncovering LLM Misalignment in RA Disease Diagnosis

   Umakanta Maharana, Sarthak Verma, Avarna Agarwal, and 4 more authors

   2025

   Under review. Investigating the implications of LLM misalignment and reasoning faithfulness in medical diagnostics.

   Abs arXiv Bib PDF Blog

   Large Language Models (LLMs) frequently achieve high classification accuracy on clinical diagnosis benchmarks while relying on medically flawed or hallucinated reasoning steps. This study evaluates diagnostic justifications generated by LLMs for Rheumatoid Arthritis (RA) against blinded clinical rheumatology experts, demonstrating that high benchmark accuracy can be deceptive and dangerous in real-world clinical deployment.

   @misc{maharana2025rightpredictionwrongreasoning,
     title = {Right Prediction, Wrong Reasoning: Uncovering LLM Misalignment in RA Disease Diagnosis},
     author = {Maharana, Umakanta and Verma, Sarthak and Agarwal, Avarna and Mruthyunjaya, Prakashini and Mahapatra, Dwarikanath and Ahmed, Sakir and Mandal, Murari},
     year = {2025},
     eprint = {2504.06581},
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2504.06581},
     note = {Under review. Investigating the implications of LLM misalignment and reasoning faithfulness in medical diagnostics.},
     keywords = {LLM misalignment, RA disease diagnosis, medical diagnostics, AI ethics, machine learning},
     journal = {arXiv preprint arXiv:2504.06581},
   }

2. EMNLP

   

   OrgAccess: A Benchmark for Role-Based Access Control in Organization-Scale LLMs

   Debdeep Sanyal, Umakanta Maharana, Hong Ming Tan, and 4 more authors

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 2025

   Accepted at EMNLP 2025. A benchmark evaluating whether models can reliably respect organizational roles, permissions, and access boundaries.

   Abs arXiv Bib PDF Blog Code Hugging Face

   We introduce OrgAccess, a benchmark for evaluating role-based access control in organization-scale LLMs, focusing on whether models can reliably respect organizational roles, permissions, and access boundaries. OrgAccess puts 16 frontier and open models against real enterprise-style access rules — 40 NIST-grounded permission types, 70K instances, up to 5 conflicting constraints at once.

   @inproceedings{sanyal2025orgaccess,
     title = {OrgAccess: A Benchmark for Role-Based Access Control in Organization-Scale LLMs},
     author = {Sanyal, Debdeep and Maharana, Umakanta and Tan, Hong Ming and Sinha, Yash and Karande, Shirish and Kankanhalli, Mohan and Kumar, Dhruv},
     booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
     year = {2025},
     eprint = {2505.19165},
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2505.19165},
     note = {Accepted at EMNLP 2025. A benchmark evaluating whether models can reliably respect organizational roles, permissions, and access boundaries.},
     keywords = {Large Language Models, Role-Based Access Control, AI Safety, Benchmark, Enterprise AI},
   }

3. EMNLP

   

   Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style

   Debdeep Sanyal, Agniva Maiti, Umakanta Maharana, and 4 more authors

   In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) , 2025

   Accepted at EMNLP 2025. An adaptive framework integrating heterogeneous student agents with genetic teacher policy evolution.

   Abs arXiv Bib PDF Blog

   Effective teaching requires adapting instructional strategies to accommodate the diverse cognitive and behavioral profiles of students, a persistent challenge in education and teacher training. We introduce a novel simulation framework that integrates LLM-based heterogeneous student agents with a self-optimizing teacher agent whose pedagogical policy dynamically evolves using a genetic algorithm. In addition, we propose Persona-RAG, a Retrieval Augmented Generation module that enables student agents to retrieve knowledge tailored to their individual learning styles.

   @inproceedings{sanyal2025pedagogical,
     title = {Investigating Pedagogical Teacher and Student LLM Agents: Genetic Adaptation Meets Retrieval Augmented Generation Across Learning Style},
     author = {Sanyal, Debdeep and Maiti, Agniva and Maharana, Umakanta and Kumar, Dhruv and Mali, Ankur and Giles, C. Lee and Mandal, Murari},
     booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
     year = {2025},
     eprint = {2505.19173},
     archiveprefix = {arXiv},
     primaryclass = {cs.AI},
     url = {https://arxiv.org/abs/2505.19173},
     note = {Accepted at EMNLP 2025. An adaptive framework integrating heterogeneous student agents with genetic teacher policy evolution.},
     keywords = {Pedagogical Agents, LLM, Multi-Agent Simulation, Retrieval-Augmented Generation, Genetic Algorithm},
   }