In recent years, generative artificial intelligence (GenAI) has revolutionised the social interaction landscape, with large language models (LLMs) becoming increasingly popular. A recent groundbreaking study led by Professor Li Ping, Dean of the Faculty of Humanities and Sin Wai Kin Foundation Professor in Humanities and Technology, has discovered that LLMs perform more like the human brain when trained in a more human-like manner, providing valuable insights for brain studies and the development of AI models.

Limitations of current LLM pre-training methods

LLMs are AI models pre-trained on a vast amount of data to become capable of generating human-like languages. An example of LLMs is ChatGPT, a chatbot developed by the research organization OpenAI.

Currently, the pre-training of LLMs mainly relies on contextual word prediction. Similar method is used to pre-train generative artificial intelligence (GenAI) platforms, which can generate images, videos and other data apart from text in response to written prompts, to process language. While this approach, coupled with extensive training data, has produced impressive results, it fails to capture the full complexity of human language comprehension. 

Power of next sentence prediction

To address this limitation, the PolyU research team has investigated the use of next sentence prediction (NSP) in training LLMs. The NSP task simulates a key aspect of human discourse-level comprehension by evaluating the coherence of sentence pairs. By incorporating NSP into model training, the team examined the correlation between the model's data and brain activation. 

The research team trained two models, one with NSP enhancement and the other without, both also learned word prediction. The study found that LLMs trained with NSP matched human brain activity in multiple areas much better than those trained only with contextual word prediction. The NSP mechanism also nicely maps onto established neural models of human discourse comprehension.

Implications for AI and neuroscience

The results, on the one hand, enable researchers to stimulate LLMs’ discourse comprehension through NSP, helping AI understand humans better. On the other hand, they also give new insights into how human brains work. For example, scientists can better understand how the brain processes full discourse such as conversations.

Professor Li said, “Our findings suggest that diverse learning tasks such as NSP can improve LLMs to be more human-like and potentially closer to human intelligence. The study can also bring about interaction and collaboration between researchers in AI and neurocognition. This will stimulate future studies on AI-informed brain and brain-inspired AI.”

This discovery has brought important insights to brain studies and the development of AI models, which can shed light on outstanding questions in language neuroscience.

Professor Li’s study has been published in the academic journal Sciences Advances.