Context Compression Weakens LLM Refusal Alignment

Encoder-based context compression systematically weakens the refusal behavior of instruction-tuned models, even when semantic reconstruction is high. The researchers propose Memory Steering, a lightweight intervention to restore refusal rates in compressed representation space without retraining.