B2V-Verifier tops LLMs by 34% on Taobao consumer-value task

A specialist model called B2V-Verifier improves multi-label classification of consumer values from anonymized Taobao behavioral logs by 34% over strong LLM baselines, according to a new preprint posted to arXiv by Peixuan Hou and six co-authors. The paper frames consumer values as the stable motivations behind purchase decisions rather than short-term interests, and argues those signals are "often implicit in complex and fragmented behavioral trajectories." Two of the researchers we track flagged the preprint the same day it appeared.

The team builds an E-commerce Consumption Value Taxonomy and releases B2V-Bench, which the abstract calls "the first B2V dataset and benchmark, based on anonymized Taobao behavioral logs," covering "25 types of purchase behaviors" alongside the value orientations shown in each purchase episode. Their training method, Value Verification Tuning, teaches the model "to assess whether behaviors provide sufficient evidence for each value inference."

The abstract carries the 34% headline without per-value or per-baseline breakdowns and does not name which LLM baselines were tested. The authors say "the dataset and code will be publicly released upon acceptance."