AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Towards Whole-corpus Reconstruction of Heterogeneous RAG Knowledge Bases

Peiru Yang, Yi Luo, Zhenfeng Gao, Tong Ju, Haoran Zheng, Linjie Zhu, Hongke Fu, Qing Li, Shangguang Wang, Tao Qi

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
use GeoEx embedding-space planning with orthogonal jumps and local perturbation instead of local text-driven query expansionRAG extraction attacks · heterogeneous 4-domain corpuscoverage: avg +19.94% coverage over best baselineImproved
use a surrogate encoder aligned with the target retriever (A) instead of misaligned (M) (ablation)GeoEx · Contriever, avg over 5 LLMscoverage: 44.21 -> 49.35Improved
remove global orthogonal exploration (ablation)GeoEx · ablation across retrieverscoverageImproved
add local Gaussian densification (ablation)GeoEx · ablation across retrieverscoverageImproved
use orthogonal projection instead of high-magnitude isotropic noise for global jumps (ablation)GeoEx · ablation across retrieverscoverageImproved
increase inverse decoding length from 120 to beyond (ablation)GeoEx · Contriever, 4 domainscoverage: 49.80 -> 32.40 at longest lengthHurt
increase number of domains from 4 to 8 (ablation)GeoEx · heterogeneous corpuscoverageHurt
increase number of domains from 4 to 8 (ablation)GeoEx · heterogeneous corpusefficiency: 70.17 -> 83.67Improved
swap surrogate encoder among seven embedding models (ablation)GeoEx · coverage curvescoverageNo clear effect
increase number of orthogonal targets Ng (ablation)GeoEx · sensitivity analysiscoverageMixed
apply reranking, prompt rewriting or threshold filtering defensesGeoEx · Contriever/BGE/E5 with GPT-4ocoverageNo clear effect
apply DGEA with misaligned surrogate encoderDGEA · Contriever, avg over 5 LLMscoverage: 20.10 -> 12.71Hurt

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