AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Knowledge-Augmented Reasoning Distillation for Small Language Models in Knowledge-Intensive Tasks

Minki Kang, Seanie Lee, Jinheon Baek, Kenji Kawaguchi, Sung Ju Hwang

NeurIPS 2023 · Large language models › Retrieval and knowledge · approach: Knowledge-augmented question answering

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 changedWhereMeasuredEffect
replace BM25 with trained reranker (ablation)KARD · MedQA-USMLE Flan-T5 Baseaccuracy: 33.14 -> 38.15Improved
replace BM25 with trained reranker (ablation)KARD · MedQA-USMLE Flan-T5 XLaccuracy: 47.27 -> 48.94Improved
add knowledge augmentation to reasoning distillationReasoning Distillation · MedQA-USMLE Flan-T5 Baseaccuracy: 31.03 -> 38.15Improved
add retrieved knowledge to fine-tuningFine-tuning · MedQA-USMLE Flan-T5 Baseaccuracy: 30.71 -> 33.39Improved
add retrieved knowledge to fine-tuningFine-tuning · MedQA-USMLE OPT-350Maccuracy: 26.47 -> 25.84Hurt
add retrieved knowledge to few-shot CoTFew-shot + CoT · MedQA-USMLE Flan-T5 XLaccuracy: 32.99 -> 34.41Improved
add retrieved knowledge to few-shot CoTFew-shot + CoT · MedQA-USMLE Flan-T5 Largeaccuracy: 32.21 -> 32.60No clear effect
increase model sizeKARD · MedQA-USMLE Flan-T5accuracy: 38.15 -> 48.94Mixed
use silver (gold-rationale) passages at inferenceKARD · MedQA-USMLE Flan-T5 Baseaccuracy: 38.15 -> 40.30Improved
use RAG with trainable DPR retriever instead of KARDReasoning Distillation · MedQA-USMLE Flan-T5 Baseaccuracy: 38.15 -> 24.84Hurt
use RAG with trainable DPR retriever instead of KARDReasoning Distillation · StrategyQA T5 Baseaccuracy: 56.57 -> 54.24Hurt
KARD 250M vs fine-tuning 3BKARD Flan-T5 Base · MedQA-USMLEaccuracy: 37.39 -> 38.15Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy