AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Enhancing LLMs for Knowledge Base Question Answering by Chain-of-Decomposition

Yonggang Zhang, Jianqi Gao, Jie Lu

ICLR 2026 · 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
decompose KBQA into trained retrieval, rule-based reformulation, and LLM reasoning (CoD)Llama2-7B · WebQSPHits@1: 84.39 -> 86.54 (vs FiDeLis GPT-4)Improved
CoD vs strongest baselineLlama2-7B · CWQHits@1: 71.47 -> 77.42Improved
CoD vs strongest baselineLlama2-7B · CWQF1: 64.32 -> 65.70Improved
use dedicated trained retriever instead of DPR/BM25/Sentence-BERTCoD · WebQSP retrievaltop5-acc: 35.02 -> 91.04 (vs DPR)Improved
swap trained retriever for DPR in CoD (ablation)CoD · WebQSPHits@1: 86.54 -> 83.37Hurt
combine logical-form executed answers and generated answers (ablation)CoD · WebQSP and CWQ (bm=5)F1: CWQ F1 49.40 / 51.37 -> 65.70Improved
increase beam size from 1 to 3 (ablation)CoD · WebQSPF1: 75.56 -> 80.87Improved
increase beam size from 3 to 5 (ablation)CoD · CWQF1: 68.27 -> 65.70Hurt
use deductive reasoning prompt instead of direct generation (ablation)CoD · WebQSPF1: 77.53 -> 81.24Improved
use deductive prompt instead of CoT prompt (ablation)CoD · WebQSPF1: 79.36 -> 81.24Improved
change backbone LLM from Llama2-7B to Qwen2.5-7B (ablation)CoD · WebQSPF1: 81.24 -> 80.39No clear effect

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