Newest papers

How Language Models Learn Taste: Brand Preference in LLMs

Tapan S. Parikh · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28272 · cited by 0 works (OpenAlex) · open-access PDF

Indonesian

Ketika diminta untuk menyebut nama merek sepatu yang berjalan, model bahasa mengatakan Nike. Diminta untuk merekomendasikan satu, mereka mengatakan Brooks. Artikel ini bertanya di mana model belajar rasa itu. Melalui 102 model dan 42 kategori merek, merek memilih model ketika merekomendasikan berbeda dari merek yang disebut di 26. Bidang bergerak bersama ke rekomendasi bersama. Merek yang disebutkan adalah pemimpin pasar, dan merek yang dipilih sebagian besar niche dan spesialis. Tiga pipa pelatihan terbuka menunjukkan di mana bentuk ini. Pretraining sudah menempelkan merek yang direkomendasikan ke saran. Pengaturan halus yang diawasi mengajarkan model untuk menjawab dan sebagian besar meninggalkan pilihan tidak berubah. Optimasi preferensi (DPO) mengajarkan model yang jawaban yang lebih baik, dan pick bergerak dalam tiga pipa, dengan 11 hingga 24 poin. Data preferensi hampir tidak mengandung merek. DPO memindahkan penilaian model ke set kedua asosi

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

Abstract

Asked to name a running shoe brand, language models say Nike. Asked to recommend one, they say Brooks. This paper asks where models learn that taste. Across 102 models and 42 brand categories, the brand a model picks when recommending differs from the brand it names in 26. The field moves together to a shared recommendation. The Named brands are market leaders, and the Picked brands are mostly niche and specialist. Three open training pipelines show where this forms. Pretraining already attaches the recommended brand to advice. Supervised fine-tuning teaches the model to answer and mostly leaves the pick unchanged. Preference optimization (DPO) teaches the model which answer is better, and the pick moves in all three pipelines, by 11 to 24 points. The preference data contains almost no brands. DPO moves the model's judgments toward a second set of associations even when the preferred brand is not in the preference data, and the effect reaches beyond brands. Asked to choose a sauce, models switch from ketchup to pesto. We do not know exactly how these preferences arise. We suspect they come from the judge models that decide which answers count as better during preference training. For brands, a recommendation depends less on what a model knows about them than on whether they fit this taste.

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

Topics: Preference, Brand preference, Advertising, Set (abstract data type), Suspect, Field (mathematics)

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