IC-566Adding instruction-specific steering vectors to the residual stream improves instruction-following accuracy for Phi-3, Gemma 2 2B IT, Mistral 7B IT, and Gemma 2 9B IT across format, length, and word-specific constraints
Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, Besmira Nushi
Steering vectors computed as the difference in residual stream activations between inputs with and without an instruction are added at a single layer during inference. Without explicit text instructions, instruction-following accuracy rises from approximately 10% to approximately 30% across all four models. With explicit instructions (baseline 60-90%), steering produces a significant boost for two of the four models. For length constraints, increasing the steering weight c continuously shortens outputs, and steering with c=20 significantly improves adherence to maximum-sentence constraints (n=1..5). For word inclusion, steering increases keyword frequency; for word exclusion, subtracting the inclusion vector reduces undesired keyword occurrence. Multi-instruction steering (format+length, casing+word exclusion) improves both constraints simultaneously. Quality scores decrease slightly under steering, comparable to the decrease from simply adding instructions as text.
Evidence
interventional
Key metric
accuracy without instructions: ~10% (no steering) to ~30% (steering); accuracy with instructions: 60%-90% (no steering), significant boost for 2 of 4 models; length: c=20, significant improvement across n=1..5 sentences (p<0.01, McNemar's test)
Caveat
Steering can occasionally produce nonsensical or repetitive tokens, particularly for 'constrained response' and 'title' instructions where no layer improves performance. Quality score decreases are small but non-zero. The method uses a fixed steering weight throughout generation and does not fully explore the intervention space.