Issue 2026-11-08 · Industry · 研究 · 应用
UNIST's EchoTrace diagnoses bias buildup in LLM recommenders

A UNIST team built EchoTrace, a tool tracking how bias and errors accumulate in LLM-based recommendation systems. It found LLMs recommended items outside the candidate set in 4.07% of movie, 7.40% of book and 13.14% of store recommendations; 93.16% of predicted user occupations were wrong.
Read original ↗