Xiaohongshu's Vela trains a 35B LLM on anonymized user behavior

Xiaohongshu's Vela turns anonymized user behavior from real-world scenarios into training signals for large language models, according to Dongxi NLP (东锡), which introduced the method.

Dongxi NLP says its 35B model improves over its base model on all 10 benchmarks, averaging +11.5 points on VelaBench and +4.0 points on general benchmarks.

The only source is Dongxi NLP's post on X, which poses the question "Can anonymized user behavior advance LLM intelligence?" and states the benchmark results without publishing the underlying scores or evaluation details.