AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

From Retrieval to Translation: Translating Query into Graph-level Clues for Retrieval-Augmented Generation

Qichuan Liu, Qinggang Zhang, Yuxuan Hu, Chenfeng Zheng, Zerui Chen, Chentao Zhang, Zhihong Zhang

ICML 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation

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
add structure-aware fine-tuning (ablation)KG-Translator (Llama) · HotpotQA/2WikiMQA/MuSiQue averageF1: 59.36 -> 68.50Improved
add structure-aware fine-tuning (ablation)KG-Translator (Llama) · 2WikiMQAF1: 63.25 -> 78.52Improved
add reverse translation (ablation)KG-Translator (Llama) · MuSiQueF1: 37.12 -> 43.32Improved
add reverse translation (ablation)KG-Translator (Llama) · average of three datasetsF1: 66.08 -> 68.50Improved
use forward-then-reverse instead of reverse-then-forward order (ablation)KG-Translator (Llama) · 2WikiMQAF1: 68.36 -> 78.52Improved
scale backbone from Llama-3.2-1B to 3BKG-Translator · MuSiQueEM: 33.89 -> 35.80Improved
scale backbone from Qwen-2.5-1.5B to 3BKG-Translator · HotpotQA and 2WikiMQAEM: 73.30 -> 72.90; 73.13 -> 72.28No clear effect
use larger embedding models instead of all-MiniLM-L6-v2 (ablation)KG-Translator · HotpotQA/2WikiMQA/MuSiQue averageF1: 69 -> 68 (0.1B), 69 (0.3B), 68 (1.5B)No clear effect
raise alpha from 1 to 3 (beta = 3) (ablation)KG-Translator · MuSiQueF1: +3.6%Improved
raise beta from 1 to 3 (alpha = 3) (ablation)KG-Translator · 2WikiMQAF1: -1.0%Hurt
increase clues from (alpha, beta) = (3, 3) to (5, 5) (ablation)KG-Translator · HotpotQA/2WikiMQA/MuSiQueF1No clear effect
replace fixed top-5 context with clue-traced retrievalKG-Translator · contexts per querycontext number: 5.00 -> 3.74Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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