arXiv:2606.00467v1 Announce Type: cross 
Abstract: Large Language Models (LLMs) are increasingly used for zero-shot annotation and LLM-as-a-judge tasks, yet their reliability hinges on how model-internalized priors interact with user-provided instructions. We investigate three dimensions of this interaction: (1) how an LLM's familiarity with data and task definitions affects performance, (2) the extent to which additional information in prompts can correct zero-shot errors ("decision stickiness"), and (3) model susceptibility to misaligned task definitions. Through experiments on toxicity detection across diverse datasets (spanning social media, gaming, news, and forums) using both dense and mixture-of-experts models, we find that nearly two-thirds of zero-shot errors are resistant to correction, with an overall rescue rate (fraction of initial errors corrected by prompting) of only 34.8%. High-confidence errors prove especially resistant to correction. When given misaligned definitions, LLMs follow them while maintaining confidence levels unchanged from the aligned condition. Crucially, we introduce Definition-Specific Familiarity (DSF), which measures alignment between a model's internal concept and the task definition. After controlling for dataset-level confounds, DSF shows a positive association with model performance (partial r = +0.41), while three distinct memorization metrics (ROUGE-L, BERTScore, and embedding cosine similarity) all fail to show a positive association. These findings show the limitations of prompt-based correction in annotation tasks, highlighting the importance of definition alignment over text-level memorization.

تسلط الأبحاث الحديثة الضوء على قيود نماذج اللغة الكبيرة (LLMs) في مهام التوصيف بدون تدريب مسبق، حيث تكشف أن ما يقرب من ثلثي الأخطاء في الكشف عن السمية مقاومة للتصحيح، مع معدل إنقاذ إجمالي منخفض يبلغ 34.8%. تفحص هذه الدراسة كيفية تفاعل الأولويات الداخلية للنموذج مع التعليمات المقدمة من المستخدم، مما يؤثر على الأداء عبر مجموعات بيانات متنوعة تشمل وسائل التواصل الاجتماعي والمنتديات.

Investigaciones recientes destacan las limitaciones de los Modelos de Lenguaje Grande (LLMs) en tareas de anotación de cero disparos, revelando que casi dos tercios de los errores en la detección de toxicidad son resistentes a la corrección, con una baja tasa de rescate general del 34.8%. Este estudio examina cómo los priors internalizados por el modelo y las instrucciones proporcionadas por el usuario interactúan, afectando el rendimiento en diversos conjuntos de datos, incluidos redes sociales y foros.

Des recherches récentes mettent en lumière les limites des grands modèles de langage (LLMs) dans les tâches d'annotation en zéro-shot, révélant que près des deux tiers des erreurs dans la détection de la toxicité sont résistantes à la correction, avec un faible taux de récupération global de 34,8 %. Cette étude examine comment les priorités internalisées par le modèle et les instructions des utilisateurs interagissent, affectant la performance sur divers ensembles de données, y compris les réseaux sociaux et les forums.

Recent research highlights the limitations of Large Language Models (LLMs) in zero-shot annotation tasks, revealing that nearly two-thirds of errors in toxicity detection are resistant to correction, with a low overall rescue rate of 34.8%. This study examines how model-internalized priors and user instructions interact, affecting performance across various datasets including social media and forums.