Research map / Large language models / Retrieval and knowledge
Thrust: Adaptively Propels Large Language Models with External Knowledge
Xinran Zhao, Hongming Zhang, Xiaoman Pan, Wenlin Yao, Dong Yu, Jianshu Chen
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 |
|---|---|---|---|
| add external knowledge only to Thrust-selected instances instead of default selection | IAPEK · e-SNLI, UnifiedQA-3b scarce | accuracy: 69.1 -> 86.3 | Improved |
| add external knowledge only to Thrust-selected instances | IAPEK · NQ, UnifiedQA-3b scarce | QA-F1: 12.8 -> 35.9 | Improved |
| add external knowledge only to Thrust-selected instances | IAPEK · TriviaQA, UnifiedQA-3b scarce | QA-F1: 39.2 -> 68.3 | Improved |
| Thrust selection on abundant portion vs scarce portion | IAPEK-Thrust · UnifiedQA-3b MC classification and QA | gain: 6.1%, 13.56% scarce vs 2.8% and 6.8% abundant | Mixed |
| Thrust selection on ARC-C | IAPEK-Thrust · ARC-C, UnifiedQA-3b scarce | accuracy: 64.5 -> 63.9 | Hurt |
| use Thrust instead of BM25 for ranking difficulty (ablation) | IAPEK · UnifiedQA-3b scarce, TREC | QA-F1: 20.4 -> 37.8 | Improved |
| use BM25 instead of Thrust (ablation) | IAPEK · Web Questions, NQ | QA-F1: 20.9 -> 19.0 | Hurt |
| vary the decoder layer used for hidden representations (ablation) | Thrust · UnifiedQA-3b, most tasks | performance | No clear effect |
| use middle-layer representations (ablation) | Thrust · StrategyQA and Web Questions | performance | Hurt |
| remove cluster size from Thrust (ablation) | Thrust · FS-NLI hardness correlation | correlation: 0.45 -> 0.23 | Hurt |
| replace vector with scalar (remove direction) (ablation) | Thrust · FS-NLI hardness correlation | correlation: 0.45 -> 0.19 | Hurt |
| remove distance (ablation) | Thrust · FS-NLI hardness correlation | correlation: 0.45 -> 0.06 | Hurt |
Closest papers by meaning
- Boosting the Potential of Large Language Models with an Intelligent Information Assistant (NeurIPS 2024 · similarity 0.87)
- Understanding the Skill Gap in Recurrent Language Models: The Role of the Gather-and-Aggregate Mechanism (ICML 2025 · similarity 0.86)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.86)
- Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers (NeurIPS 2025 · similarity 0.86)
- From Static to Dynamic: Knowledge Metabolism for Large Language Models (AAAI 2024 · similarity 0.86)
- Self-Updatable Large Language Models by Integrating Context into Model Parameters (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.
