Implicit Reasoning for Large Language Model-based Generative Recommendation

In three linesPauseRec introduces an implicit reasoning paradigm for LLM-based generative recommendation systems. The approach bypasses explicit reasoning pipeline limitations by avoiding reasoning trace acquisition, improving performance by up to 6.22%, reducing GPU training costs by 65%, and accelerating inference by 71.3%.

Summary generated by Claude — human-verified