Language-Switching Triggers Take a Latent Detour Through Language Models

Researchers traced a Latin trigger in an 8B autoregressive language model that flips English responses into French.

The paper finds a three-stage circuit: early attention heads assemble the trigger, a latent subspace carries it, and the final-layer MLP converts it into French output.

This shows how a hidden backdoor can work below the surface, bypassing simple input checks and evading common defenses.

The finding highlights a new model-security risk in internal representations and language-switching attacks.