IC-567Steering vectors computed on instruction-tuned Gemma 2 models transfer to base Gemma 2 models, with cross-model steering outperforming same-model steering for Gemma 2 2B

Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, Besmira Nushi

SourceImproving Instruction-Following in Language Models through Activation Steering

The paper computes steering vectors for format instructions on the instruction-tuned versions of Gemma 2 2B and 9B, then applies them to the corresponding base models. In both the no-instruction and with-instruction settings, cross-model steering yields performance gains over no steering. For Gemma 2 2B specifically, cross-model steering outperforms same-model steering (vectors computed on the base model itself) across both settings. For length instructions, a conciseness vector computed on Gemma 2 9B IT, when applied to the base model, shortens outputs as the steering weight increases. The authors note this is the first demonstration that cross-model steering can outperform same-model steering in base models.

Evidence
interventional
Caveat
Experiments are limited to the same model family and parameter size (Gemma 2 2B and 9B, base vs. instruction-tuned). Transferability across different architectures or sizes is not tested.
Model
Gemma 2
Concepts
Linear representation
Datasets
IFEval / IFEval-Simple [eval]
Methods
Activation steering / Mean steering / PCA steering [primary]
Related work
Ilharco et al. (task arithmetic) [context], Kissane et al. (SAEs usually transfer between base and chat models) [context]
Related findings
IC-565, IC-566, IC-568
Extraction
automatic-extraction