The new probe architecture can aggregate information across multiple layers and multiple token positions, whereas traditional linear probes are usually trained only on a single layer or the last token.