60 semantically equivalent references with diverse expressions. Further, LLMs are widely employed as the evaluators of text generation in a reference-free manner, including evaluating a single prediction [631, 632, 642] or comparing several candidates [138, 643645]. Nevertheless, LLMs may expose bias (e.g., order bias or preference for LLM-generated texts over human-written texts) as language generation evaluators, demonstrating disparities when compared to human evaluation [632, 646, 647]. Unreliable Generation Evaluation LLMs have been capable of generating texts with a comparable quality to human-written texts, which however might be underestimated by automatic reference-based metrics. As an alternative evaluation approach, LLMs can serve as language generation evaluators to evaluate a single text, compare multiple candidates, and improve existing metrics. However, this evaluation approach still needs more inspections and examinations in real-world tasks.

Underperforming specialized generation. Although LLMs have learned general language patterns to generate coherent text, their proficiency in generation might be constrained when dealing with a specialized domain or task. For instance, a language model that has been trained on general web articles may face challenges when generating a medical report which involves many medical jargon and methods. Intuitively, domain knowledge should be critical for model specialization. However, it is not easy to inject such specialized knowledge into LLMs. As discussed in recent analyses [47, 648], when LLMs are trained to exhibit some specific ability that allows them to excel in some areas, they might struggle in others. Such an issue is related to catastrophic forgetting [649, 650] in training neural networks, which refers to the conflict phenomenon of integrating new and old knowledge. Similar cases also occur in human alignment of LLMs, where alignment tax (e.g., a potential loss in the

Moreover, due to the limitations of sequence modeling architecture, LLMs still face challenges in the understanding and generation of structured data. Consequently, they often fall behind task-specific models on complex structured data tasks, such as knowledge-base question answering and semantic parsing [458, 651]. Therefore, it is important to develop effective model specialization methods that can flexibly adapt LLMs to various task scenarios, meanwhile retaining the original abilities as possible. Underperforming Specialized Generation LLMs may fall short in mastering generation tasks that require domain-specific knowledge or generating structured data. It is non-trivial to inject specialized knowledge into LLMs, meanwhile maintaining the original abilities of LLMs.

7.1.2 Knowledge Utilization Knowledge utilization is an important ability of intelligent systems to accomplish knowledge-intensive tasks (e.g., commonsense question answering and fact completion) based on supporting factual evidence. Concretely, it requires LLMs to properly utilize the rich factual knowledge from the pretraining corpus or retrieve external data when necessary. In particular, question answering (QA) and knowledge completion have been two commonly used tasks for evaluating this ability. According to the test tasks (question answering or knowledge completion) and evaluation settings (with or without external resources), we categorize existing knowledge utilization tasks into three types, namely closed-book QA, open-book QA43, and knowledge completion.