Research map / Large language models / Retrieval and knowledge
Beyond Perplexity: Let the Reader Select Retrieval Summaries via Spectrum Projection Score
Zhanghao Hu, Qinglin Zhu, Siya Qi, Yulan He, Hanqi Yan, Lin Gui
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| rank sampled summaries by SPS instead of PPL (ablation) | xCompress (Llama 3.1 8b) · NQ | EM: 35.0 -> 39.4 | Improved |
| rank by SPS instead of PPL (ablation) | xCompress (Qwen3 8b) · HotpotQA and NQ | EM: 29.6 -> 28.8 (HotpotQA); 30.0 -> 28.0 (NQ) | Hurt |
| add SPS-based test-time sampling to CompAct | CompAct (Llama 3.1 8b) · NQ | EM/F1: 35.2/47.49 -> 39.4/51.18 | Improved |
| add SPS selection to xRAG | xRAG (Mistral-7b) · TriviaQA | EM: 16.0 -> 29.2 | Improved |
| max pooling instead of mean pooling for SPS embedding (ablation) | SPS (Llama 3.1) · NQ | EM: 37.2 -> 39.6 | Improved |
| max pooling instead of last-token pooling (ablation) | SPS (Llama 3.1) · Musique | EM: 5.8 -> 9.2 | Improved |
| use penultimate layer embeddings and 0.95 PCA variance (ablation) | SPS · QA benchmarks | EM/F1 | Improved |
| generate more than five candidate summaries (ablation) | xCompress · TriviaQA | EM: EM 65.4 at five summaries | No clear effect |
| score summaries with PPL or LongPPL instead of SPS (correlation with answer quality) (ablation) | summary scoring · HotpotQA | PCC (EM): 0.643 (SPS) vs 0.022 (PPL) | Hurt |
Closest papers by meaning
- SePer: Measure Retrieval Utility Through The Lens Of Semantic Perplexity Reduction (ICLR 2025 · similarity 0.90)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.90)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.89)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.89)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.89)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
