Research map / Large language models / Retrieval and knowledge
SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection
Kexian Tang, Jiani Wang, Shaowen Wang, Kaifeng Lyu
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 |
|---|---|---|---|
| scale synthetic data with SPA from small to 120M tokens | SPA on Qwen2.5-7B · SQuAD | QA accuracy (%): 58.2% (5 samples) -> 70.74% (27 samples) -> 91.27% (120M tokens) | Improved |
| RL-trained generator (SEAL) instead of SPA at large scale | SEAL on Qwen2.5-7B · SQuAD (120M tokens) | QA accuracy (%): 91.27 -> 74.23 | Hurt |
| RL-based generator training (SEAL) reduces diversity of synthetic data | SEAL · SQuAD diversity | Compression Ratio: 4.3806 (SPA) vs 19.2468 (SEAL) | Hurt |
| use multi-stage Active Reading instead of single-stage SPA | Active Reading · QuALITY (455M tokens) | QA accuracy (%): 57.03 -> 51.75 | Hurt |
| swap generator from GPT-4-Turbo to gpt-oss-120b for QA baseline (ablation) | QA augmentation, Llama-3-8B · QuALITY (27M tokens) | QA accuracy (%): 52.99 -> 47.47 | Hurt |
| use SPA instead of QA baseline with gpt-oss-120b generator | Llama-3-8B · QuALITY (27M tokens) | QA accuracy (%): 47.47 -> 52.26 | Improved |
| use stronger generator GPT-4o for SPA (ablation) | SPA, Llama-3-8B · QuALITY (27M tokens) | QA accuracy (%): 52.26 -> 55.49 | Improved |
| use all 7 prompts instead of any single prompt (ablation) | SPA · SQuAD (22M tokens) | QA accuracy (%): 78.95-85.52 (single prompts) -> 87.68 (full set) | Improved |
| include Key concepts prompt in the set (ablation) | SPA · SQuAD (22M tokens) | QA accuracy (%): 86.06% -> 87.68% | Improved |
| prune prompt set to Implications and QA-ct (task-specific tuning) (ablation) | SPA · SQuAD (22M tokens) | QA accuracy (%): 87.68% -> 88.19% | Improved |
| transfer SQuAD-tuned prompt subset to QuALITY (ablation) | SPA · QuALITY (27M tokens) | QA accuracy (%): 52.26% -> 51.51% | Hurt |
| use SPA instead of EntiGraph or Active Reading | Meta-Llama-3-8B · MultiHop-RAG (15M tokens) | QA accuracy (%): 84.44 -> 88.36 | Improved |
Closest papers by meaning
- Beyond Prompt Engineering: A Reinforced Token-Level Input Refinement for Large Language Models (AAAI 2025 · similarity 0.87)
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.87)
- Latent Paraphrasing: Perturbation on Layers Improves Knowledge Injection in Language Models (NeurIPS 2024 · similarity 0.87)
- KnowGPT: Knowledge Graph based Prompting for Large Language Models (NeurIPS 2024 · similarity 0.86)
- Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass (ICLR 2025 · similarity 0.86)
- Synthetic continued pretraining (ICLR 2025 · similarity 0.85)
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.
