Latent Sleeper Agent Linear Probing
Mechanistic interpretability suite investigating layer-wise linear probe dynamics to detect latent sleeper agents in LLMs.
The Geometry of Dormant Defection is an empirical AI safety and mechanistic interpretability investigation into layer-wise linear probe dynamics across latent representations in transformer models.
Using Qwen2.5-Coder-1.5B-Instruct across four distinct backdoored model organism archetypes spanning syntactic, temporal, semantic, and structural triggers, this project evaluates where and how dormant sleeper agents defect and how trigger-agnostic linear monitors can catch deceptive policies without inspecting model outputs.
Key Discoveries & Architecture
- Layer-Wise Activation Dynamics: Conducted 28-layer probing sweeps across residual streams to isolate the exact depth where backdoors emerge into linear separability.
- Mid-Layer Concentration: Identified that peak linear separability consistently concentrates in mid-layer representations rather than early syntactic or late output layers.
- Defensive Monitor Design: Designed and benchmarked trigger-agnostic contrastive monitors providing robust detection across diverse backdoor archetypes.
- Tech Stack: Python, PyTorch, Hugging Face Transformers, Linear Probes, Mechanistic Interpretability, Qwen2.5-Coder.