Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Score

This paper introduces Spectrum Projection Score (SPS), a metric that measures how well retrieved information aligns with a reader language model’s internal representations without requiring labelled data. Building on SPS, we propose xCompress, a framework that selectively samples and ranks retrieval summaries to improve retrieval-augmented generation across multiple question-answering benchmarks.

Recommended citation: Z Hu, Q Zhu, S Qi, Y He, H Yan, L Gui. (2025). "Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Score." Proceedings of the AAAI Conference on Artificial Intelligence (AAAI 2026).
Download Paper