IC-166A 1-dimensional subspace in a single layer encodes the context-versus-prior decision in Llama-3.1-8B, Gemma-2 9B, and Mistral-v0.3 7B, and setting this subspace steers the released (non-fine-tuned) models' behavior
Julian Minder, Kevin Du, Niklas Stoehr, Giovanni Monea, Chris Wendler, Robert West, Ryan Cotterell
The authors identify a rank-1 subspace in layer 16 (Llama-3.1) of the residual stream that encodes whether the model will follow the context or its prior knowledge. Although the subspace is learned on a fine-tuned model, the same direction functions as an effective control knob in the non-fine-tuned instruct and base models of each family. Steering the subspace value (e.g., c(pri)=-6, c(ctx)=6 for Llama-3.1) changes the model's answer without any explicit intent instruction in the prompt. A strong Pearson correlation between the subspace value separation and task performance (0.908 for Llama-3.1, 0.871 for Mistral, 0.975 for Gemma-2) further supports that this subspace is the mechanism underlying context sensitivity.
Evidence
interventional
Key metric
Llama-3.1-8B-Instruct zero-shot: 73% pairacc baseline, 73% steered; Llama-3.1-8B ICL: 83% baseline, 88% steered; Gemma-2 9B instruct zero-shot: 4% baseline, 44% steered; Pearson correlation between subspace value mean difference and pairacc: 0.908 (Llama-3.1, p=0.000), 0.871 (Mistral, p=0.000), 0.975 (Gemma-2, p=0.000)
Caveat
Steering is less effective when the model must reason over the context rather than directly copy the answer; the base model zero-shot achieves 0% pairacc because it lacks instruction-following training; the subspace was learned on the fine-tuned model and its transfer to released models is not uniform across all configurations.