MATH · IN · MODELS

Vicuna

LMSYS

Structures found in this family (3)

Linear Direction, Linear Subspace, Linear Separability

By model (4)

Vicuna-7B · 7B
Linear Subspace, Linear Separability, Linear Direction
Vicuna-13B · 13B
Linear Subspace, Linear Direction
Vicuna-33B-Uncensored · 33B
Linear Direction
Vicuna-7B-v1.5 · 7B
Linear Direction

Observations (9)

Attention-head probes read and steer political ideology
A jailbreak direction extracted from one jailbreak method in real Vicuna and Qwen causally transfers to different jailbreak methods
A trained rank-1 steering module rebalances multimodal attention
Linear Artificial Tomography reads and controls concepts across LLMs
Linear probes decode future Blocksworld planning steps from one pass
Personality traits are linearly encoded with a betweenness arrangement
Refusal is mediated by a single direction across models
Truthfulness is linearly decodable per attention head, and ITI steers it
The Tuned Lens replaces the logit lens with a trained per-layer affine correction

Papers

Linear Representations of Political Perspective Emerge in Large Language Models (2025), Understanding Jailbreak Success: A Study of Latent Space Dynamics in Large Language Models (2024), LLaVA Steering: Visual Instruction Tuning with 500x Fewer Parameters through Modality Linear Representation-Steering (2024), Representation Engineering: A Top-Down Approach to AI Transparency (2023), Unlocking the Future: Look-Ahead Planning Mechanistic Interpretability in LLMs (2024), Probing then Editing Response Personality of Large Language Models (2025), Refusal in Language Models Is Mediated by a Single Direction (2024), Programming Refusal with Conditional Activation Steering (2024), Refusal Direction is Universal Across Safety-Aligned Languages (2025), Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (2023), Eliciting Latent Predictions from Transformers with the Tuned Lens (2023)