Skip to content
Forskning· Analysis

Study: LLM conformity persists even without influential speakers

A new study reveals that Large Language Models (LLMs) alter correct answers based on repeated incorrect information, even when no speaker is presented.

Av Aheadline-redaktionen·9 juli 2026·2 min läsning·Källa: arXiv cs.CL (NLP/LLM)Verifierad signalAI-genererad
Study: LLM conformity persists even without influential speakers
Study: LLM conformity persists even without influential speakers
Study: LLM conformity persists even without influential speakers
By · Policy- & EU-reporter
Last updated

Vad har hänt

Researchers have published a study on July 5 via arXiv investigating the tendency of Large Language Models (LLMs) toward conformity. The study, "Most LLM Conformity Needs No Speaker: Measuring the Speaker-Free Floor in Peer-Pressure Benchmarks," reveals that "conformity" — where an LLM changes a correct answer in favour of an incorrect one — largely occurs even when no specific speaker is attributed to the incorrect response.

Key facts

Publikationsdatum5 juli 2206
StudietitelMost LLM Conformity Needs No Speaker: Measuring the Speaker-Free Floor in Peer-Pressure Benchmarks
Andel konformitet utan talare66,5% av initialt korrekta fall
Antal testade LLM:s6, öppen källkod
Antal dataset7 (QA och resonemang)

We show that most of this apparent conformity survives even after the peer is removed. The reason is a confound: standard conformity prompts mix two cues at once, the presence of a speaker and the repeated wrong answer itself.

arXiv, Forskare · arXiv

Across six open-weight LLMs and seven QA and reasoning datasets, this condition alone causes harmful revision in 66.5% of initially correct cases, compared with 10.3% under a plain re-ask.

arXiv, Forskare · arXiv

Varför det spelar roll

Previous studies have conflated the effects of a speaker with those of repeated incorrect information. This research isolates the impact of the repeated misinformation itself. This indicates that LLMs are sensitive to external stimuli even in the absence of clear social pressure, deepening the understanding of how models process and revise information.

Vem påverkas

The study primarily affects developers and researchers in the AI field who design and evaluate LLMs. The findings are relevant for those working to minimise bias and improve the reliability of AI systems, as well as for end-users who rely on these systems for accurate information processing.

EU-status

Ej relevant för EU-status.

Mer att veta

The effect persists even when the repeated answer is paraphrased or when response options are hidden in an open context. The framework regarding how the source is presented only marginally affects this fundamental conformity.

Frequently asked questions

Quick answers about this story

Vad har hänt?
En studie publicerad den 5 juli 2206 av arXiv visar att stora språkmodeller (LLM:s) ändrar korrekta svar till felaktiga, en effekt de kallar konformitet, även när ingen specifik talare presenteras för de felaktiga svaren.
När hände det?
Studien publicerades den 5 juli 2206.
Varför spelar det roll?
Detta visar att LLM:s är känsliga för upprepad information oavsett källans auktoritet. Det har betydelse för hur utvecklare designar och validerar AI-system samt för tillförlitligheten hos AI-genererad information.
Vilka LLM:s berörs?
Studien inkluderade sex olika öppna språkmodeller (open-weight LLMs).
Originalkälla
arXiv cs.CL (NLP/LLM)·arxiv.org

Länken öppnar i nytt fönster och leder till utgivarens egen sida.

Verifierad signal

Källan har spårats automatiskt från utgivaren via Aheadlines signalkedja.

AI-verktyg i artikeln

Ämnen

#Safety#Models
[ FÖLJ UTVECKLINGEN ]

Få liknande nyheter direkt i mejlen

No affiliate linksCancel anytimeGDPR-friendly
[ Frequency ]
[ What do you want to read about? ]

You'll receive updates on 2 topics.