[2311.07468] An Analysis and Mitigation of the Reversal Curse

View a PDF of the paper titled An Evaluation and Mitigation of the Reversal Curse, by Ang Lv and Kaiyi Zhang and Shufang Xie and Quan Tu and Yuhan Chen and Ji-Rong Wen and Rui Yan

View PDF
HTML (experimental)

Summary:Latest analysis noticed a noteworthy phenomenon in giant language fashions (LLMs), known as the “reversal curse.” The reversal curse is that when coping with two entities, denoted as $a$ and $b$, linked by their relation $R$ and its inverse $R^{-1}$, LLMs excel in dealing with sequences within the type of “$aRb$,” however encounter challenges when processing “$bR^{-1}a$,” whether or not in technology or comprehension. For example, GPT-4 can precisely reply to the question “Tom Cruise’s mom is?” with “Mary Lee Pfeiffer,” however it struggles to supply a passable reply when requested “Mary Lee Pfeiffer’s son is?” On this paper, we undertake the first-ever research of how the reversal curse occurs in LLMs. Our investigations reveal that the reversal curse can stem from the particular coaching goals, which change into notably evident within the widespread use of next-token prediction inside most causal language fashions. We hope this preliminary investigation can draw extra consideration to the reversal curse, in addition to different underlying limitations in present LLMs.