Djamé Seddah

Detecting a backdoor trigger does not mean you can stop it: SAE features in attention and MLPs reliably flag trojan inputs but lack causal control, whereas sparse directions in the residual stream hold full steering power over backdoor execution.

LLMs may ace academic trivia but falter on everyday cultural knowledge, revealing a critical gap in their understanding.

Implicit multi-hop reasoning in LLMs is fundamentally constrained by pretraining exposure, not knowledge gaps.

Turns out, you can boost a multilingual LLM's reasoning in any language just by swapping in the mid-layers from an English-reasoning specialist.