We released Aligner: a new efficient alignment paradigm, bypasses the whole RLHF process. 无需RLHF显著提升GPT-4/Llama2性能,北大团队提出Aligner对齐新范式