Newest papers

Where Do LLMs Learn to Recommend AI?

Benaya Trabelsi, Jonathan Shaki, Sarit Kraus · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.29785 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Preguntado qué estudiar, qué carrera seguir, o qué producto construir, los grandes modelos de lenguaje (LLMs) a menudo recomiendan opciones relacionadas con la IA. Nos preguntamos de dónde viene esta tendencia comparando modelos antes y después del ajuste fino y variando la forma en que se piden las preguntas. En primer lugar, los modelos pre-entrenados ya incluyen términos explícitos de IA/ML en las recomendaciones del 27% de sus respuestas, y sus versiones pre-entrenadas publicadas lo hacen en 59%. En segundo lugar, los modelos de ajuste fino nos muestran que este aumento no es inevitable: el ajuste general de instrucciones deja la tasa sin cambios, pero la eliminación del 7% de los ejemplos de instrucciones que se preocupan por la IA lo corta por 14 puntos porcentuales, por debajo del

Machine translation (Meta M2M100); the English original follows.

Abstract

Asked what to study, which career to pursue, or what product to build, large language models (LLMs) often recommend AI-related options. We ask where this tendency comes from by comparing models before and after fine-tuning and by varying how questions are asked. First, pretrained models already include explicit AI/ML terms in the recommendations of 27% of their responses, and their released instruction-tuned versions do so in 59%. Second, fine-tuning models ourselves shows that this rise is not inevitable: general instruction tuning leaves the rate unchanged, but removing the 7% of instruction examples that concern AI cuts it by 14 percentage points, below the level of the pretrained models it starts from. Data rewarding helpfulness raise it by 9 points, while data ranked by the preferences of diverse simulated users lower it by 9. Finally, with models held fixed, asking for the "best" option or for what to do "right now" raises the rate by 8 and 18 points, whereas telling the model it is an AI advisor adds 4. Overall, the tendency starts in pretraining, and what models learn afterward, together with how a question is phrased, decides how strongly it appears.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Helpfulness, Ask price, Product (mathematics), Computer science, Psychology, Language model

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange