Does context compression preserve refusal alignment?

The study finds that encoder-based context compression, while maintaining high reconstruction fidelity, systematically weakens the refusal behavior of instruction-tuned language models, and this effect persists across different model families and compression architectures. Mechanism analysis shows that compression attenuates activations along the refusal direction learned by the decoder. The authors further propose Memory Steering, a lightweight inference-time intervention that operates entirely within the compressed representation space and restores refusal rates to near-baseline levels without retraining. The results indicate that semantic preservation does not guarantee behavioral preservation under compression, and compression-aware systems need to explicitly preserve alignment-related features.