CS graduate student
Manning College of Information and Computer Sciences
University of Massachusetts Amherst
Email: sagnikchatte@umass.edu
Research interests
Mechanistic interpretability and AI agents. I work on reverse-engineering model internals — circuits, features, and attention patterns — and on using what that reveals to build systems that reason, plan, and act more reliably. I am particularly drawn to how capabilities emerge in transformers (superposition, polysemanticity, in-context learning), and to what a clearer mechanistic picture implies for the design of tool-using agents.
Writing
- A Practical Survey of Mechanistic Interpretability Techniques
- Reproducing CRISP: a flat line where the paper has a 25-point drop
- Can Protein Language Models Tell Which Genes Work Together? Operon Classification with Siamese Networks
- Does Showing Your Work Help LLMs Learn Tools? A Continual Learning Experiment
- My Teeny Tiny Understanding of Diffusion Models
Projects
- Decodability is not localization — tracing math errors in Qwen2.5-Math
Probing Qwen2.5-Math-1.5B on ProcessBench shows a linear probe decodes "this trace has gone wrong" at AUROC 0.866, yet names the first wrong step in only 28.9% of cases and fires on 39% of fully correct solutions — a study of the gap between ranking and thresholded localization.
- Kai
a minimal, hackable coding assistant that runs entirely in the terminal with local models through Ollama.
- Refusal Steering with Sparse Autoencoders
testing when sparse-autoencoder features are worth their cost over simple linear directions for detecting and steering refusal in language models.
- Fine-Tuning Qwen2.5 with LoRA and Unsloth
a small, end-to-end fine-tuning pipeline for a 4-bit Qwen2.5 chat model, built to understand what each part of LoRA training actually does.
Elsewhere
GitHub | LinkedIn | Reading list | Resume