Cross-Architecture Steering Transfer in Language Models: A Systematic Empirical Study

The paper presents the first systematic evaluation of cross-model steering transfer, testing whether concept directions learned from one LLM can be used to control a different, independently trained model. The authors train one sparse autoencoder per model across 15 semantic domains, covering five open-weight models from two architectural lineages at scales from 0.8B to 8B parameters, and evaluate all 20 directed model pairs.

They find a clear scale dependence: for models with at least 1.7B parameters, 47-49% of cross-model feature pairs validate (Pearson r >= 0.60, Procrustes cosines 0.895-0.956), while alignment degrades sharply below 0.8B. Cross-model steering vectors achieve a 71.0% win rate across 15 supervised concepts, outperforming same-model native vectors (68.0%). Remarkably, a single universal steering vector works across 4 of 5 models without per-model supervision, achieving 67.3% accuracy.

Transfer fails for models below 1.7B parameters and for one model with generation instability, indicating that functional exploitability requires sufficient representational capacity. The findings establish a scale threshold for mechanistic interpretability tools and provide a functional complement to the Platonic Representation Hypothesis, showing that geometric convergence across models can support cross-model behavioral control without fine-tuning.