LLMs can rank job candidates more efficiently and accurately by processing them as a list rather than individually, but require special techniques to overcome position bias and the 'lost-in-the-middle' problem.
This paper addresses position bias and inefficiency in LLM-based talent recommendation systems. The authors propose L3TR, a listwise ranking framework that processes multiple candidates together rather than one-by-one, using block attention and positional encoding techniques to reduce token consumption and improve recommendation quality.