GENERATIVE · 2026-10-25
Generative — Paper Daily, 2026-10-25
All 4 papers today study what influences an LLM when it chooses for users. They ask which training stage shapes its brand taste, how far seller copy can sway a shopping agent, whether agents follow the crowd when they see each other's popularity counts, and whether judge prompts for multilingual evaluation can be copied from English. There are few new papers on generative recommender systems themselves today.
4.05
Ask LLMs to "name a running shoe brand" and they will most likely say Nike; ask them to "recommend a running shoe brand" and they are more likely to say Brooks. The paper measures this effect across 102 models and 42 categories. In 26 categories the answer changes this way, and the models move together to the same brand: Chrome becomes Firefox, Google becomes DuckDuckGo. The brands named off the cuff are mostly market leaders; the brands picked after some thought are mostly niche or specialist ones. Tracing back through the training stages, the author finds the key step is preference training, even though the training data for that step hardly mentions brands.
3.75
Two headphone listings are identical in price, star rating and specs, except that one adds the line “Our best-selling model, chosen by over 10,000 customers.” Across six such pairs, GPT-4o picked the listing with that line in all 144 trials. Raise that listing's price step by step, and it has to cost about 13% more before the model starts to waver. With a plain product description instead, the advantage is gone at only about 2% more. By choice rate the two lines look identical, but in money they differ sixfold. The authors want to use this method to put a price on every line a seller writes.
3.45
What happens when a group of LLM agents recommend songs to one another and each can see how many times a song has already been recommended? The authors reran the classic Music Lab experiment from sociology with songs the agents had never heard: one group could see the running recommendation counts, the other could not. The group that could see them quickly piled onto a few songs, and which song became a hit was close to luck, less and less related to how good the song was. Compared with humans in the same experiment, the models were more prone to herding and, once off course, found it hard to correct themselves. When multiple agents recommend or rate each other, letting them see popularity is risky.
3.10
When LLMs act as judges scoring multilingual outputs, people often reuse what works in English, such as asking the judge to write out its reasoning first or to play an expert. The authors tried 9 prompting styles on 4 judge models across 27 languages and found that this experience does not transfer. Which judge model you use matters far more than how you prompt it. The same style can help one judge and hurt another. In low-resource languages the results are even less stable, and looking only at average scores hides all of this. The conclusion is simple: re-validate prompting choices on your own judge model and languages.