mech-interp

Pivotal Token Search

Adversarial Manipulation of CoT

Analysed determinism, faithfulness, reasoning patterns, & steering. Developed and tested methods to enhance control and fail-safes

mech-interp suite for Granite4 models that use Mamba-2 architecture

All code, stimuli, and results for a mechanistic interpretability study investigating how large language models internally represent emotional content

Implementation and analysis of Sparse Autoencoders for neural network interpretability research. Features interactive visualization dashboard and W&B integration.

Detect safety degradation during LLM fine-tuning before it becomes behavioral

Collection and learnings of my journey in Artificial Intelligence

Official implementation of the 'Uncovering Competency Gaps in Large Language Models and Their Benchmarks' paper

Investigated JEPA-WM predicted futures, specifically how activation steering can impact robotics task planning and outcomes. Performed a series of ablations to test hypotheses about CEM action planning in latent space for world models.

Local agent-driven mechanistic interpretability research platform for Apple Silicon

Unofficial implementation to reproduce the experiments from "Superposition as a Phase Change" of "Toy Models of Superposition".

A re-implementation of the "Refusal in Language Models is Meditated by a Single Direction" to understand mech interp

Cross-architecture mechanistic interpretability toolkit — first OSS Mamba SSM state extraction. Works on transformer + SSM + hybrid models with unified API.