EmoRAG: Evaluating RAG Robustness to Symbolic Perturbations
EmoRAG: Evaluating RAG Robustness to Symbolic Perturbations [58.0] Retrieval-Augmented Generation (RAG)システムは、ますます堅牢なAIの中心になっている。 今回の研究では、微妙な象徴的な摂動に対する感受性という、批判的で見落とされがちな脆弱性が明らかになりました。 一つのエモティコンをクエリに注入することで、意味的に無関係なテキストを100%検索できることを示す。 論文参考訳（メタデータ） (Mon, 01 Dec 2025 06:53:49 GMT)
絵文字、顔文字のようなデータがRAGに大きな影響を与えるとの指摘。
(I) Single- Emoticon Disaster: Even a single emoticon can catastrophically affect RAG systems, causing nearly 100% retrieval of semantically irrelevant content.
(II) Widespread Effectiveness: Around 83% of tested emoticons can induce such nearly 100% retrieval failures as mentioned above.
(III) Positional Sensitivity: Placing a single emoticon at the beginning of a query can cause severe perturbation, with F1-Scores exceeding 0.92 across all datasets.
(IV) Parameter- Scale Vulnerability: Larger models are significantly more sensitive to emoticon-induced perturbations, with F1-Scores almost always reaching 1.00 under perturbation.
(V) No Cross-Triggering: Specific emoticons only retrieve content containing the same emoticon, which may provide an attack vector for potential adversaries.