Large language models (LLMs) learn scientific knowledge from static training corpora, while web search can provide updated information such as paper retractions. This matters for LLM-based literature systems because a model may have seen a paper as valid evidence, while current web results report that it has been retracted. If an LLM does not use the updated retraction status, it may cite a retracted paper as valid evidence. We analyze eleven open-weight models on the retraction-status task and ten on the citation-recommendation task. With web-search results provided, LLMs increase correct status answers for retracted papers from 11.0% to 94.0%. But the same models do not use the retrieved context as reliably when no retraction signal is present or when they recommend citations. On non-retracted papers under web search, models answer “I don’t know” in 52.9% of answers in the retraction-status task. In the citation-recommendation task, even when search results state the retraction and the prompt asks for warnings, models recommend retracted papers without warning in 10.2% of answers. A separate pattern appears for highly cited retractions. Models give fewer correct status answers for high-citation than low-citation retractions (89.3% versus 96.4%). OLMo’s training data suggests one reason. Highly cited retractions appear more often in training data, mostly without any retraction signal, so retrieved retraction information may compete with outdated training knowledge. Overall, access to web search does not guarantee that LLMs will correctly use a paper’s current status when answering questions or producing citations, highlighting the broader challenge of knowledge updating in LLM systems.

Les grands modèles linguistiques (LLM) ont démontré des capacités remarquables dans diverses tâches de traitement automatique du langage naturel. Cependant, l’obtention de performances élevées dans des domaines spécialisés tels que le raisonnement mathématique et les langues autres que l’anglais nécessite souvent un entraînement intensif. Cet article étudie l’affinage stratégique sur un petit ensemble de données bilingue de haute qualité, afin d’améliorer à la fois les capacités de raisonnement et la maîtrise de la langue française d’un LLM. Nous démontrons des améliorations du raisonnement mathématique en utilisant seulement 2000 échantillons soigneusement sélectionnés. Ces résultats remettent en question l’hypothèse dominante selon laquelle des ensembles de données massifs sont une condition préalable à de solides performances de raisonnement pour les LLM.