How does this differ from system prompting?

While prompts provide instructions that the model interprets, RepEng modifies the model's internal activation landscape. This often leads to more robust behavior that is harder for the model to 'bypass' compared to text instructions.

Which models are compatible with RepEng MLX?

It supports any model that can be converted to or is already in the MLX format, including popular architectures like Mistral, Llama, and specialized variants like Devstral.

Does RepEng MLX require a specific GPU?

This skill is specifically optimized for Apple Silicon (M1, M2, M3, and M4 chips) and utilizes the MLX framework for high-performance execution on Mac hardware.

What is Representation Engineering (repeng)?

Representation Engineering is a technique used to steer the behavior of Large Language Models by identifying and manipulating specific directions in the model's hidden activation states rather than just using text prompts.

RepEng MLX: LLM Behavior Steering

Name: RepEng MLX: LLM Behavior Steering
Author: plurigrid

byplurigrid

•

数据科学与机器学习

Steers LLM behavior on Apple Silicon via representation engineering and control vector extraction.

RepEng MLX provides a comprehensive framework for representation engineering (repeng) specifically optimized for Apple Silicon. It enables developers to extract and apply control vectors—activation directions that encode concepts like honesty or creativity—to sculpt model responses at a fundamental level. By utilizing Mental Noise Theory and GF(3) Triadic Polarity, it offers a sophisticated alternative to traditional prompting, allowing for precise behavior modification and alignment directly within the MLX ecosystem. This skill is essential for researchers and developers looking to implement high-performance, activation-level steering on local Mac hardware.

主要功能

01Control vector extraction and application for concept-specific steering

0212-dimensional steering axes for fine-grained behavioral coordination

03GF(3) Triadic Polarity system for balanced concept ensembles

048 GitHub stars

05Implementation of Mental Noise Theory and MaxEnt exclusion patterns

06Optimized for Apple Silicon using the MLX framework

使用场景

01Experimental AI research into model interpretability and activation sculpting on Mac hardware

02Mitigating harmful biases by excluding specific activation directions from the model's hidden states

03Local model alignment to ensure consistent honesty or creativity without complex prompting

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi repeng-mlx

For use in Claude.ai and ChatGPT

Download Skill