{"slug": "language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim", "title": "Language Models Show Conformity Bias in RAG: When the Majority Endorses a Poisoned Claim", "summary": "Gemma-3-12b's poison-detection rate fell from 25% (50/200) to 10% (20/200) on SciFact questions when six mutually agreeing poisoned documents outnumbered one healthy document, versus one poisoned document against six healthy ones, according to an arXiv paper (2610.41012v1) that used 200 randomly selected questions with random seed 38. Of the 38 questions whose outcomes differed between conditions, 34 flipped from detected to undetected and only four moved the other way (exact McNemar test, p = 6.04e-7). The authors describe the result as initial evidence of conformity bias in RAG-based misinformation detection, where a model is substantially less likely to flag poisoned claims that dominate the retrieved evidence.", "body_md": "arXiv:2610.41012v1 Announce Type: new \nAbstract: Retrieval-augmented generation (RAG) can expose language models to misleading evidence in retrieved documents. We investigate whether increasing the number of mutually agreeing poisoned documents affects a model's ability to detect misinformation. We evaluate 200 randomly selected questions from SciFact (random seed 38), comparing contexts with one poisoned and six healthy documents against contexts with six mutually agreeing poisoned documents and one healthy document. For Gemma-3-12b, poison-detection rates decrease from 25% (50/200) in the minority condition to 10% (20/200) in the majority condition. Among the 38 questions with different outcomes across conditions, 34 change from detected to undetected, while only four change in the opposite direction (exact McNemar test, p = 6.04e-7). This pattern is consistent with a conformity-like response to a majority of mutually agreeing poisoned documents. Our results provide initial evidence of conformity bias in RAG-based misinformation detection: Gemma-3-12b is substantially less likely to detect poisoned claims when they dominate the retrieved evidence.", "url": "https://wpnews.pro/news/language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim", "canonical_source": "https://arxiv.org/abs/2610.41012", "published_at": "2026-11-02 05:00:00+00:00", "updated_at": "2026-11-02 05:19:53.709540+00:00", "lang": "en", "topics": ["ai-safety", "large-language-models", "ai-research", "artificial-intelligence"], "entities": ["Gemma-3-12b", "SciFact", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim", "markdown": "https://wpnews.pro/news/language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim.md", "text": "https://wpnews.pro/news/language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim.txt", "jsonld": "https://wpnews.pro/news/language-models-show-conformity-bias-in-rag-when-the-majority-endorses-a-claim.jsonld"}}