jasperfu/activation-steering-demo

Activation Steering — Live Demo

Edit what a language model is "thinking" by adding a concept vector to a middle-layer activation. No fine-tuning, no prompt engineering.

What is this

Based on Anthropic's July 2026 paper Verbalizable Representations Form a Global Workspace in Language Models, which shows that LLMs maintain a privileged set of internal representations ("workspace") that can be read via a Jacobian lens and directly manipulated via activation patching.

This Space demonstrates the manipulation side: pick a direction (via positive/negative word contrast), and edit any middle-layer activation with h_ℓ ← h_ℓ + α · direction.

Try

1. 1.Pick a preset (Sentiment / Formality / Style / Verbosity)
2. 2.Edit the positive/negative words, or write your own from scratch
3. 3.Adjust α (0–100) and see the three columns spread from opposite direction ← baseline → steered direction

Sweet spot α is typically 20–60 for the 0.6B model.

Model

Qwen/Qwen3-0.6B — chosen because it fits on the free CPU tier. For stronger effects you'd want 4B+.

What works, what doesn't

• —✅ Sentiment on review-style prompts
• —✅ Formality on communication tasks
• —⚠️ Aesthetic on generative tasks (partial)
• —❌ Specific typography / long-form style consistency
• —❌ Prompts with strong factual anchors (product specs)

Recipe

direction = mean(embed([" amazing", " excellent", ...])) - mean(embed([" terrible", " awful", ...]))

def hook(module, input, output):
    h = output[0]
    h[:, :] += alpha * direction
    return (h,) + output[1:]

model.model.layers[layer - 1].register_forward_hook(hook)
model.generate(...)   # first forward gets steered