AI Paper Atlas

Research map / Vision-language and multimodal / Visual question answering

EntRAG: Entity-Centric Retrieval-Augmented Generation for Knowledge-based Visual Question Answering

Yiheng Hu, Xiaoyang Wang, Qing Liu, Xiwei Xu, Qian Fu, Wenjie Zhang

ICML 2026 · Vision-language and multimodal › Visual question answering · approach: Attention and prior-robust VQA

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
use entity-conditioned visual processor instead of CLIP visual encoder (ablation)EntBind · E-VQA entity retrievalR@1: 42.4 -> 57.0Improved
use entity-guided attention pooling instead of learned-weight pooling (ablation)EntBind · E-VQA entity retrievalR@1: 53.4 -> 57.0Improved
add visual embedding (visual processor) (ablation)EntBind · E-VQA entity retrievalR@1: 21.0 -> 57.0Improved
add entity token embedding (ablation)EntBind · E-VQA entity retrievalR@1: 51.4 -> 57.0Improved
combine similarity and relevance scores for entity ranking (ablation)ECRanker · E-VQA entity-levelR@1: 62.8 -> 78.0Improved
rank contexts within top-ranked entity (Entity-First) instead of pooled contexts (ablation)ECRanker · E-VQA context-levelR@1: 59.4 -> 68.0Improved
use joint score instead of Entity-First for context ranking (ablation)ECRanker · E-VQA context-levelR@1: 68.0 -> 62.3Hurt
Entity-Pool instead of Entity-First ranking (ablation)ECRanker · E-VQA context-levelR@5: 77.4 -> 81.2 (R@5), 68.0 -> 59.4 (R@1)Mixed
use EntBind alone for context-level ranking (ablation)EntBind · E-VQA context-levelR@1: 68.0 -> 6.8Hurt
use CAGen with oracle context (ablation)LLaVA-v1.6-7B · E-VQA oracle retrievalVQA accuracy: 78.1 -> 88.7Improved
entity-centric retrieval with EntBind instead of MM-EmbedEntRAG · InfoSeek entity retrievalR@1: 36.7 -> 58.5Improved
swap generator backbone from LLaVA-v1.5 to Qwen-2.5EntRAG · InfoSeek (All)VQA accuracy: 40.3 -> 44.5Improved

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