IC-149Intervening in the shared representation space using the dominant language (English) predictably changes model outputs for other data types, demonstrating the space is causally used rather than a vestigial byproduct
The paper performs activation addition (ActAdd) and activation replacement interventions in intermediate layers, using English trigger words, on models processing non-English or non-textual inputs. For multilingual steering, ActAdd with English sentiment words ('good'/'bad') on Llama-3 processing Spanish or Chinese text achieves the intended sentiment shift with minimal fluency loss, comparable to steering in the input language. For code, adding the unembedding difference of 'start'-'end' at layer 17 steers Llama-2 to expand range(end) into range(0,end) in up to 67% of cases. For vision, replacing image patch hidden states with the unembedding of a different color word steers Chameleon to report the new color in >80% of cases. For audio, adding the unembedding difference of a non-mammal word to Salmonn's hidden states shifts the mammal prediction probability in a dose-dependent manner. For arithmetic, ActAdd with English number words steers Llama-2 and Llama-3 to output a decremented number followed by '+1' in >90% of cases.
Evidence
interventional
Key metric
Llama-3 Spanish down-steering with English: sentiment 0.139±0.026 vs baseline 0.143±0.022; Llama-2 code range() steering: up to 67% correct at max intervention strength; Chameleon color replacement: >80% success rate at optimal layer; Salmonn mammal steering: dose-dependent shift in p(mammal) from ~0.575 to ~0.725; arithmetic intervention: >90% of instances steered correctly (n=846)
Caveat
The authors note that interventions are not guaranteed to produce sensible outputs, only the intended directional change. For Chameleon, more sophisticated reasoning tasks (e.g., country flags, spatial relationships) are beyond the model's capability even without intervention. The code intervention uses a simpler method than ActAdd (direct unembedding difference addition). The color replacement is more invasive than addition and may lexically bias generation, though the two-color control partially addresses this.