Bio

Hi, I am Niyar R Barman, an M2 student at Université de Toulouse, working with Dr. Nicholas Asher.

I work on reasoning in language models, mechanistic interpretability, and efficient systems for training and analysis. I have papers at EMNLP 2023 (outstanding paper) and SIGDIAL 2025.

My research asks when models are actually reasoning and when they are relying on patterns that look right on benchmarks but fail under distribution shift. I study this by building structured reasoning setups, running mechanistic experiments to locate and measure shortcut behavior inside networks, and testing hypotheses with causal interventions, targeted ablations, and out-of-distribution evaluations.

More recently, I have been working closer to the systems stack, writing CUDA and Triton kernels to understand how model computation behaves at the hardware level and how training and inference can be made more efficient without collapsing useful internal representations (the log).

Open to research and industry roles after October 2026.

You can reach me at: niyar-r.barman[at]utoulouse[.]fr

News

  • Apr 2026: Joined the R&D Team at (LINAGORA Labs) for M2 Internship
  • Apr 2026: TELL-TALE: Task Efficient LLMs with Task Aware Layer Elimination accepted at the Findings of ACL 2026
  • Sep 2025: Started M.S. in Interactions of Computer Science and Mathematics for AI at Université de Toulouse, France
  • Aug 2025: DIMSUM: Discourse in Mathematical Reasoning as a Supervision Module accepted at SIGDIAL 2025
  • Jan 2025: Returned to IRIT, Toulouse for a second full-time research internship
  • May 2024: Started first full-time research internship at IRIT, Toulouse, France
  • Dec 2023: Counter Turing Test CT² presented at EMNLP 2023, received Outstanding Paper Award
  • Jan 2023: Started remote research internship at the Artificial Intelligence Institute, University of South Carolina, USA
  • Dec 2021: Started B.Tech. in Electronics and Communication Engineering at National Institute of Technology Silchar, India

Publications

selected · all
ACL 2026 Findings
TELL-TALE: Task Efficient LLMs with Task Aware Layer Elimination
Omar Naim, Krish Sharma, Niyar R Barman, Nicholas Asher
TALE boosts LLM inference by removing task-irrelevant layers to match or exceed baseline accuracy across 9 tasks and 5 model families while reducing compute, requiring no retraining and only modest setup cost.
paper · code
SIGDIAL 2025
DIMSUM: Discourse in Mathematical Reasoning as a Supervision Module
Niyar R Barman, Krish Sharma, Nicholas Asher, Akshay Chaturvedi
Proceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue
This work explores how discourse structures can be used as supervision for improving mathematical reasoning in large language models.
paper · code
EMNLP 2023
outstanding paper award
Counter Turing Test CT²: AI-Generated Text Detection is Not as Easy as You May Think
Megha Chakraborty, S. M. Towhidul Islam Tonmoy, S. M. Mehedi Zaman, Krish Sharma, Niyar R Barman, Chandan Gupta, Shreya Gautam, Tanay Kumar, Vinija Jain, Aman Chadha, Amit P. Sheth, Amitava Das
Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
Introduces the AI Detectability Index and challenges current approaches to AI-generated text detection.
paper
Preprint 2024
The Brittleness of Image Watermarking Techniques: Investigating Visual Paraphrasing for De-Watermarking AI-Generated Images
Niyar R Barman, Krish Sharma, Ashhar Aziz, Shashwat Bajpai, Shwetangshu Biswas, Aman Chadha, Vasu Sharma, Amitava Das
Demonstrates that current AI image watermarking methods are fragile and can be circumvented through visual paraphrase attacks using image captioning and diffusion systems.
paper
IEEE CICT 2023
A Transformer-Based Approach to Automate Disease Prediction from Patient Descriptions
Niyar R Barman, Krish Sharma, Ranjay Hazra
IEEE 7th Conference on Information and Communication Technology
Proposes a transformer-based approach for automated disease prediction using textual symptom descriptions to provide accurate diagnoses.
paper
FIRE 2023
Addressing Hate Speech: ATLANTIS for Efficient Hate Span Detection
Niyar R Barman, Krish Sharma, Yashraj Poddar, Adwaitha Vatupal, Partha Pakray
FIRE 2023 Working Notes
Presents a hate span detection system capable of identifying consecutive sets of hateful content in text.
paper

Research Experience

  • May 2024 - Jul 2024, Jan 2025 - Mar 2025, Research Intern, Institut de Recherche en Informatique de Toulouse, France
  • Jan 2023 - Dec 2024, Research Intern, Artificial Intelligence Institute of University of South Carolina, United States

Education

  • 2025 - 2026, M2 Master Interactions de l’Informatique et des Mathématiques pour l’IA, Université de Toulouse, France
  • 2021 - 2025, Bachelors of Technology in Electronics and Communication Engineering, National Institute of Technology Silchar, India