Peer-reviewed research
Six papers across LLM interpretability, model compression, and AI-generated text detection, with two as first author. Each entry links to the paper or code.
CircuitTier: Multi-Signal Interpretability for Efficient Model Compression
A task-aware way to quantize large language models. Instead of compressing every component equally, CircuitTier fuses six interpretability signals to set per-component bit-widths, protecting the parts that matter for the task. On Gemma-2-2B for SQL generation it keeps about 99% of task accuracy at 3.84x compression, and the result holds across architectures, with 3.84x at 97.6% retention on Llama-3-8B.
Disentangling Direction and Magnitude in Transformer Representations
An interpretability study that separates the direction and the magnitude of Transformer representations. Using L2-matched perturbations, it demonstrates a double dissociation between the two, clarifying which part actually carries task information inside the model.
Fine-Grained Detection of AI-Generated Text Using Sentence-Level Segmentation
Detecting AI-generated text at the sentence level rather than labeling a whole document, which allows much more precise attribution of which parts of a text were machine-written.
CNLP-NITS-PP at GenAI Detection Task 1: AI-Generated Text Using Transformer-Based Approaches
A transformer-based system for the shared task on detecting machine-generated text.
nits_teja_srikar at GenAI Detection Task 2: Distinguishing Human and AI-Generated Essays
Machine learning and transformer models to tell human-written essays apart from AI-generated ones.
CNLP-NITS-PP at GenAI Detection Task 3: Cross-Domain Machine-Generated Text Detection Using DistilBERT
A DistilBERT-based approach for detecting machine-generated text that generalizes across domains.