ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs | Read Paper on Bytez