GENERATIVE · 2026-10-25
Generative — Paper Daily, 2026-10-25
Today’s four papers all ask how LLMs make choices for users: where brand recommendation taste comes from, how much seller copy can steer shopping agents, whether multi-agent popularity feedback herds recommendations, and whether multilingual judge prompts transfer from English. Fresh generative-rec system papers are scarce in this announcement window; the set leans toward auditing and evaluating LLM recommenders and shopping agents.
4.05
Ask a model to name a running-shoe brand and it says Nike; ask it to recommend one and it says Brooks. Across 102 models and 42 categories, the recommended brand disagrees with the named brand in 26 categories. Named brands are mostly market leaders; picked brands are mostly niche specialists. Across three open training pipelines, pretraining already attaches the recommended brand to advice-style prompts, SFT mostly leaves the pick unchanged, and DPO moves it by 11–24 points in every pipeline. The preference data contains almost no brand comparisons, yet DPO still shifts judgments—even sauce choices from ketchup to pesto. Recommendation taste forms mainly in preference optimization, not only from what the model “knows” about brands.
3.75
When an LLM shops for a user, seller cues such as “best-seller,” endorsements, or scarcity can steer its pick. Choice-rate audits hit a ceiling once a cue already wins near 100% of the time. Price-equivalent (PE) auditing raises the cued product’s price in a counterbalanced forced choice until it is chosen only half the time, and reports that premium as the cue’s PE. Across 113,472 trials, seven models, and two categories, cues that look equally strong by choice rate can differ several-fold in PE. The decision-native model Jev resists most cues yet still bends to the strongest social-proof claims. Pricing seller content in money makes different cues comparable.
3.45
When LLM agents recommend unfamiliar songs and can see cumulative recommendation counts, they form herding cascades like the classic Music Lab. Seeing counts makes collective picks diverge from a song’s intrinsic appeal into extreme, unpredictable outcomes. LLM populations herd more readily than humans and correct themselves less. Visible popularity among agents amplifies feedback loops instead of converging on intrinsic quality.
3.10
Multilingual LLM-as-a-Judge setups often reuse prompting tricks validated only in English. This study runs 9 prompting strategies, 4 open-weight judges, 27 languages, and 4 tasks across 13 datasets for >1M pointwise scores against gold references, checked with human judgments. Prompting strategies do not transfer: the judge model dominates the prompt choice, the same strategy can reverse across judges, and stability drops in low-resource languages. Validate prompts for the intended judge, task, and language instead of copying English defaults.