Despite their advanced reasoning capabilities, the latest LLMs often miss the mark when deciphering relationships. In this article, we explore the Reversal Curse, a pitfall that affects LLMs across tasks such as comprehension and generation.
Large Language Models (LLMs) are renowned for their advanced reasoning capabilities, enabling them to perform a wide range of tasks, from natural language processing to code generation. However, despite their strengths, LLMs often exhibit a weakness in deciphering relationships, particularly when dealing with inverses. This phenomenon, termed the “reversal curse,” affects LLMs across various tasks, including comprehension and generation.
To understand the underlying issue, let’s consider a scenario with two entities, denoted as a and b, connected by their relation R and its inverse. LLMs excel at handling sequences such as “aRb,” where a is related to b by relation R. For instance, an LLM can quickly answer the question, “Who is the mother of Tom Cruise?” when asked. However, LLMs struggle with the inverse relation, denoted as R inverse. In our example, if we ask an LLM, “Who is Mary Lee Pfeiffer’s son?” it is more likely to hallucinate and falter, despite already knowing the relationship between Tom Cruise and Mary Lee Pfeiffer.
This reversal curse is a pitfall that affects LLMs in a variety of tasks. In a recent study, researchers from the Renmin University of China brought this phenomenon to the attention of the research community, shedding light on its probable causes and suggesting potential mitigation strategies. They identify the Training Objective Function as one of the key factors influencing the extent of the reversal curse.
To fully grasp the reversal curse, we must first understand the training process of LLMs. Next-token prediction (NTP) is the dominant pre-training objective for current large language models, such as GPT and Llama. In models like GPT and Llama, the attention masks during training depend on the preceding tokens, meaning each token focuses solely on its prior context. This makes it impossible to account for subsequent tokens. As a result, if a occurs before b in the training corpus, the model maximizes the probability of b given a over the likelihood of a given b. Therefore, there is no guarantee that LLMs can provide a high probability for a when presented with b. In contrast, GLM models are pre-trained with autoregressive blank in-filling objectives, where the masked token controls both preceding and succeeding tokens, making them more robust to the reversal curse.
The authors put this hypothesis to the test by fine-tuning GLMs on “Name to Description” data, using fictitious names and feeding descriptions to retrieve information about the entities. The GLMs achieved approximately 80% accuracy on this task, while Llama’s accuracy was 0%.
To address this issue, the authors propose a method that adapts the training objective of LLMs to something similar to ABI. They fine-tuned models using Bidirectional Causal Language Model Optimization (BICO) to reverse-engineer mathematical tasks and translation problems. BICO adopts an autoregressive blank infilling objective, similar to GLM, but with tailored modifications designed explicitly for causal language models. The authors introduced rotary (relative) position embeddings and modified the attention function to make it bidirectional. This fine-tuning method improved the model’s accuracy in reverse translation and mathematical problem-solving tasks.
In conclusion, the authors analyze the reversal curse and propose a fine-tuning strategy to mitigate this pitfall. By adopting a causal language model with an ABI-like objective, this study sheds light on the reversal underperformance of LLMs. This work could be further expanded to examine the impact of advanced techniques, such as RLHF, on the reversal curse.
Don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. If you like our work, you will love our newsletter. Don’t Forget to join our 55k+ ML SubReddit.
