Alpaca

Stanford

Structures found in this family (2)

Linear Subspace, Linear Direction

By model (2)

Alpaca-7B · 7B

Linear Subspace, Linear Direction

Alpaca · 7B

Linear Direction

Observations (3)

Boundless DAS scales causal-alignment search to a 7B LLM

Human evaluation confirms per-layer style vectors steer emotion

Truthfulness is linearly decodable per attention head, and ITI steers it

Papers

Interpretability at Scale: Identifying Causal Mechanisms in Alpaca (2023), The Effectiveness of Style Vectors for Steering LLMs: A Human Evaluation (2026), Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (2023)