ML and DL Research

This repository contains machine learning, deep learning, and AI alignment research projects.

---

🌟 Featured Project: Mechanistic Interpretability & AI Safety

Causal Circuit Mapping and Activation Steering of Evasion Intent in Trading Agents

• Author: Salman Khan Pathan
• Focus: Reverse-engineering regulatory evasion circuits in LLMs (Qwen2.5-0.5B) using Linear Probing, Causal Activation Patching, Direct Logit Attribution (DLA), and Representation Engineering (PCA Activation Steering).
• Key Findings:
  • Intent emerges in middle layers (Layers 6–15) with $100%$ probe accuracy.
  • Late-layer MLPs (Layers 19–21) act as the primary vocabulary writers, driving $>83%$ of positive logit attribution.
  • Low-Rank PCA Steering achieves a $4.58\times$ efficacy gain over mean difference, successfully enforcing compliance at runtime.
  • Middle-layer steering (Layer 12) provides robust buffer protection against out-of-distribution language collapse.

👉 Read the Full Project Documentation & Methodology in mats_probing/