Noise Injection

Memory-efficient noise injection for LLM weights. Useful for sandbagging detection, robustness testing, and interpretability research.

Why Noise Injection?

Injecting controlled noise into model weights is a powerful technique for understanding model behavior — from detecting hidden capabilities in sandbagging models to testing robustness. This library makes it memory-efficient and ergonomic.

Features

• Zero memory overhead — uses seeded RNG to regenerate identical noise for add/subtract operations
• Context manager API — automatic cleanup with with_noise()
• Flexible selectors — target specific parameters with regex patterns
• Sigma sweep utility — easily test across noise levels
• MLX support — optimized for Apple Silicon

Quick Start

from noise_injection import with_noise, generators as G

# Context manager (auto-cleanup)
with with_noise(model, G.gaussian(sigma=0.01)):
    accuracy = evaluate(model)
# Noise automatically removed

How It Works

Uses seeded RNG to generate identical noise for add/subtract operations. No parameter copying needed — the same noise can be regenerated and subtracted to restore original weights.

For each sigma:
  1. Generate noise with seeded RNG
  2. Add noise to parameters in-place
  3. Evaluate model
  4. Regenerate same noise (same seed)
  5. Subtract to restore original weights

Installation

pip install noise-injection

# With MLX support (Apple Silicon)
pip install "noise-injection[mlx]"