---
title: "Discussion on \"Directional Ablation Can't see 24° Rotation of the Refusal Direction\" | Hashnode"
description: "Discussion on \"Directional Ablation Can't see 24° Rotation of the Refusal Direction\". Why directional ablation fails to detect refusal vector rotation after fine-tuning, and how calibration decouples estimate drift from safety."
url: "https://hashnode.com/posts/directional-ablation-can-t-see-24-rotation-of-the-refusal-direction/6ad0ab8ff68738517b9e43ae"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Directional Ablation Can't see 24° Rotation of the Refusal Direction](https://mechanistic-artifacts.hashnode.dev/directional-ablation-can-t-see-24-rotation-of-the-refusal-direction?utm_source=hashnode&utm_medium=feed)

Why directional ablation fails to detect refusal vector rotation after fine-tuning, and how calibration decouples estimate drift from safety.

Oct 15 · 23 min read · mechanistic-artifacts.hashnode.dev  
Read the article

-   [#mechanistic-interpretability](https://hashnode.com/tag/mechanistic-interpretability)
-   [#ai-safety-and-alignment](https://hashnode.com/tag/ai-safety-and-alignment)
-   [#post-training](https://hashnode.com/tag/post-training)
-   [#llms](https://hashnode.com/tag/llms)
-   [#fine-tuning](https://hashnode.com/tag/fine-tuning)

## Written by

[Viraj Sawant](https://hashnode.com/@viraaj)

@viraaj · Mumbai

  
[View profile](https://hashnode.com/@viraaj)

Research. Experiment. Build. Repeat

I'm an AI Engineer and Researcher passionate about understanding, building, and advancing intelligent systems. My work spans Large Language Models (LLMs), reinforcement learning, deep learning, computer vision, and AI engineering. I enjoy exploring how models learn, designing experiments, investigating research questions, and translating ideas from papers into practical implementations. Here, I share research insights, technical deep dives, experiments, implementation guides, and lessons learned while building AI systems. \* My goal: Bridge the gap between AI research and real-world engineering through rigorous experimentation, reproducible work, and open knowledge sharing.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.