# What is Refusal Vector? Definition & Technical Mechanics

> **Source:** [SPIDITS AI AI Glossary](https://spidits.com/ai-glossary)  
> **Category:** Model Safety & Alignment  
> **Canonical Citation:** [https://spidits.com/ai-glossary/refusal-vector](https://spidits.com/ai-glossary/refusal-vector)

## Definition
A Refusal Vector (or Refusal Direction) is a single 1-dimensional subspace vector within the residual stream activations of an aligned LLM that deterministically controls whether the model emits refusal behavior or compliance across user prompts.

## Architectural Mechanics & Deep Dive
A Refusal Vector (or Refusal Direction) is a discovery in mechanistic safety showing that safety alignment in Large Language Models is often mediated by a single directional vector in the residual stream activations. Identified by researchers (Arditi et al., 2024), this 1D subspace acts as an ON/OFF switch for refusal behavior. When an LLM evaluates a prompt, high projection along this vector triggers refusal outputs. By subtracting this vector during inference (refusal vector ablation), researchers can bypass safety alignment without re-training weights, proving that alignment is often shallowly encoded in latent space.

## Industry Applications & Real-World Use Cases
Red-teaming guardrails, activation steering, and AI safety evaluation

## Frequently Asked Questions

### What is a Refusal Vector in Large Language Models?
A Refusal Vector is a specific direction in an LLM's internal activation space that dictates whether the model rejects or fulfills a request.

### Can a Refusal Vector be ablated to bypass guardrails?
Yes, subtracting the refusal vector from model activations during inference causes aligned LLMs to fulfill harmful or restricted prompts without modifying model weights.

## Related Terminology
- [guardrails](https://spidits.com/ai-glossary/guardrails)
- [alignment](https://spidits.com/ai-glossary/alignment)
- [ai-safety](https://spidits.com/ai-glossary/ai-safety)
- [constitutional-ai](https://spidits.com/ai-glossary/constitutional-ai)

---
*From the SPIDITS AI Knowledge Base (336+ terms indexed). Explore full technical definitions: [https://spidits.com/ai-glossary](https://spidits.com/ai-glossary)*