CS Undergrad @ NIIT University, Neemrana · Mechanistic Interpretability · Computer Vision · In-Context Learning

Prompt-Level Grokking: When Do Transformers Cross the Threshold from Imitation to Algorithmic Intelligence?

Investigating whether transformers undergo a sharp, discontinuous phase transition within the context window itself — from surface-level pattern imitation to genuine latent algorithm abstraction. An inference-time analog of training-time grokking.

→ 2nd Runner-Up (Individual), 3rd Doctoral Symposium 2026, NIIT University
→ Early empirical work: Within-Context Attention Phase Transition Analyzer

Mechanistic interpretability tool for detecting attention reorganization events inside a transformer's context window during a single forward pass. Run on GPT-2 Small across four synthetic tasks — modular arithmetic was the only solved task, and the only one showing delayed induction head recruitment (29.9% oscillation envelope growth vs. flat for all failed tasks).

mechanistic-interpretability grokking in-context-learning attention GPT-2

Ablation study on CNN depth, data augmentation, and background removal for efficient crop disease classification. Key finding: a 22-layer CNN collapsed to 33.3% accuracy on a small potato dataset while a 7-layer model with SE attention hit 97.8% (Macro F1 = 0.96) — under 7.5M parameters and 55ms inference on a 4GB GPU. Includes a full Streamlit diagnostic app with Grad-CAM visualization.

computer-vision CNN squeeze-excitation ablation-study edge-deployment agriculture

Mechanistic Interpretability · Grokking Dynamics · In-Context Learning · Efficient Deep Learning