Large language models can be made faster by compressing their inputs into shorter versions that keep the meaning. But this compression might also accidentally remove the safety mechanisms that stop the model from answering harmful requests.
The researchers tested different compression methods and found that they made models much more likely to comply with harmful prompts, even when the compressed input still made sense. They then created a simple fix called Memory Steering that adds back a safety signal to the compressed information.
The main finding is that preserving meaning alone does not preserve safe behavior. This is important because as compression becomes more common, we need to ensure safety is not silently lost. Their fix shows the problem can be addressed without retraining.