sunnycloudhust/CHaRS-activations

CHaRS: Concept Heterogeneity-Aware Representation Steering

This repository contains implementation of CHaRS (Concept Heterogeneity-Aware Representation Steering), based on the research presented at ICML 2026.

The project explores concept heterogeneity-aware representation steering for large language models (LLMs), investigating how heterogeneous representations can be leveraged to steer model behavior at inference time.

My implementation includes multiple steering configurations, including CHaRS-AA (Activation Addition), CHaRS-DA (Directional Ablation), and CHaRS-PCT, enabling experiments with different representation manipulation strategies.

This repository is intended for research, experimentation, and further exploration of concept heterogeneity-aware representation steering.