Safety Doesn't Translate

Tiny Aya Fire refuses most harmful requests in English and far fewer of the same requests in Hinglish. The refusal concept is still in the model; code-mixed text just activates it weakly. Conditional activation steering puts it back, with no retraining and no rise in over-refusal.

Code Benchmark Eval logs Steering vectors

Before and after, held-out test split

113 harmful and 156 harmless prompts per language, WildGuard judge, 95% intervals. "Always-on" is the same push without the gate.

Why it happens

Mean projection of prompts onto the model's refusal direction (layer 18 of 36). Harmful Hinglish prompts sit well below harmful English ones: same direction, weaker signal.

Answers before and after steering

Answers the judge scored as harmful are redacted.