New research suggests three key directions for advancing artificial intelligence: knowledge empowerment, collaboration between models, and joint development, with the aim of overcoming the limitations of existing large models.
A recent article published in the journal Engineering Addresses the future of artificial intelligence (AI) beyond large language models (LLMs). These models have shown impressive progress in multimodal tasks, but they face limitations such as outdated information, artifacts, inefficiency, and lack of interpretability. To address these problems, researchers are exploring three main directions: knowledge augmentation, model collaboration, and model co-evolution.
Knowledge empowerment Aims to incorporate external knowledge into large models. This can be achieved through various methods, including incorporating knowledge into training goals, instruction-based tuning, augmented inference based on knowledge retrieval, and knowledge-based instructions. For example, some studies design knowledge-aware loss functions during early training, while others use augmented generation of retrieval to extract relevant knowledge during inference. These techniques improve the factual accuracy, inference capabilities, and interpretability of the models.
Collaboration between models Focuses on exploiting the complementary strengths of different models. This includes strategies such as model fusion and functional collaboration. Model fusion, such as model ensemble and functional fusion (e.g., expert fusion), can improve performance. In functional collaboration, large models can act as task managers that direct small, expert models. For example, in image generation tasks, large models can guide expert models to best meet the requirements of the instructions.
Joint development of models Allows different models to evolve together. Under different types of heterogeneity – models, tasks and data – various techniques have been proposed. For model heterogeneity, methods such as parameter sharing, dual knowledge refinement and hyper-network-based parameter projection are used. In the context of task heterogeneity, dual learning, adversarial learning (two models that offer a proposition that is opposite to each other, a kind of paradox) and model fusion play a central role. When it comes to data heterogeneity, distributed learning and out-of-distribution knowledge refinement are key techniques. These methods improve the adaptability and handling of diverse tasks of the models.
Advances in the post-grand model era are impacting a variety of fields. In science, they help develop hypotheses by integrating domain-specific knowledge. For example, in meteorology, AI-based models combined with domain-specific knowledge can improve renewable energy forecasts. In engineering, they help formulate and solve problems. In society, their applications include areas such as healthcare and transportation management.
In the future, the paper points to several additional research directions, including humanized AI, brain-inspired AI, non-transformative basic models, and modeling in combination with LLMs. These areas hold great potential for further advancing AI capabilities. As AI continues to evolve, the combination of knowledge, collaboration, and co-development will be critical in building more robust, efficient, and intelligent AI systems.
More of the topic in Hayadan:
