Substantial efforts have been made to investigate the source of emergent abilities. Scaling law [14,15] is a great representative of these works, which reveals that scaling can greatly improve model capacity and eventually lead to emergent abilities. Kaplan et al. [14] firstly explore the relationship between the performsnce of LLMs and important factors. By fitting the model performance with varied data sizes (from 22 million to 23 billion tokens), model sizes (from 768 million to 1.5 billion non-embedding parameters), and training compute, they found that the model performance has a strong dependence relation on model sizes, data sizes, and training compute, which can be formulated as:

Later, Hoffmann et al. [15] proposed an alternative form of scaling law. Differently, they conducted rigorous experiments by varying a larger range of model sizes (from 70 M to 16 B) and data sizes (from 5 B to 500 B tokens), and fitted a similar scaling law yet with different coefficients:

The experiment showcased that model size makes more contribution to emergence abilities than does data size. In summary, scaling law reveals that as the number of parameters and data size increase, the model can represent more abstract language features, thereby achieving generalization to new tasks and obtaining emergence ability.

In addition, a myriad of factors during the training and application of LLMs have an impact on emergent capabilities. For example, the diversity and quality of training data have been shown to significantly influence emergent abilities [16]. Diverse data cover various patterns, domains, and contexts, which enable LLMs to learn general representations and provide the foundations for emergent abilities. High-quality datasets offer rich task-related context, which allows LLMs to learn different task patterns, enhances the generalization of LLMs on unseen tasks, and facilitates the emergent of LLMs. In addition to training dataset, diverse pretraining tasks are able to improve generalization of LLMs and foster the emergence of emergent abilities [7]. Specifically, instruction tuning provides comprehensive task descriptions for LLMs by designing and collecting instruction data covering multiple tasks. During the optimization phase, LLMs obtain accurate understanding of tasks, enabling LLMs to comprehend a wide range of tasks and exhibit emergent capabilities. In downstream tasks, different prompt formats, which are generated through manual [17] or automated optimization [18], tend to have different impacts on emergent abilities. As research advances, Schaeffer et al. [19] argued that LLMs do not have the so-called emergence ability and the phenomena of emergent abilities are just an illusion in the process of evaluating the performance of LLMs. By using a range of metric to evaluate the capabilities of LLMs, they found that the emergent abilities of LLMs are observed under the condition of using nonlinear or discrete metrics. In contrast, when using linear or continuous metrics, the curve of LLMs’ performance becomes smooth, continuous, and predictable. Consequently, the phenomena of emergent abilities are due to the selection of metrics rather than fundamental changes in model behavior with scaling model size and data size. Moreover, they conduct meta-analysis on open-sourced benchmarks to further prove their hypotheses. Finally, the authors extended their experiments to visual tasks, including image reconstruction and image classification. The findings suggest that emergent abilities are not unique to LLMs but rather a widespread issue inherent to evaluation methodologies.

Based on reinforcement learning, reinforcement learning from human feedback (RLHF) [20] regards human feedback as a reward signal to guide LLMs in generating human-aligned responses. Once proposed, RLHF quickly becomes a popular framework of human alignment (Fig. 4). However, RLHF depends heavily on high-quality human feedback, which is both costly and time-consuming due to the need for qualified annotators. To mitigate this issue, reinforcement learning from AI feedback (RLAIF) [21] was introduced. In contrast to RLHF, RLAIF employs AI agents to generate response–preference pairs rather than human annotators, which reduces the demand for human annotators and enables large-scale feedback generation.

In addition to designing feedback strategies, researchers have paid attention to the impact of reinforcement learning techniques on human alignment. Given that RLHF is a complex and often unstable process, Rafailov et al. [22] proposed the direct preference optimization (DPO) algorithm, taking advantage of the mapping relationship between the reward function and the optimal policy. It demonstrates that this constrained reward maximization problem is essentially a classification problem solved on human preference data and can be precisely optimized through single-stage policy training. To improve the sampling efficiency of proximal policy optimization (PPO), Liang et al. [23] designed PTR-PPO, which combines both on-policy and off-policy methods. In practice, by introducing prioritized trajectory replay, PTR-PPO effectively reuses past experiences and improves sampling efficiency. Different from traditional RLHF approaches, which is computationally expensive, SuperHF [24] introduces a supervised fine-tuning paradigm, leading to a more data-efficient and computationally feasible human alignment. Existing human preference alignment methods assume that the preference relationships between different choices are monotonic and transitive, providing a reasonable approximation of human preferences. However, human decision-making is effected by various factors, leading to inconsistency and nonlinearity. To address this, Wu et al. [25] introduced the self-play framework SPPO. In this framework, LLMs are fine-tuned based on their own previous rounds, optimizing through model-generated synthetic data and annotations from the preference model.

In early work, the “retrieval-then-read” framework [26] was widely adopted. In this framework, a retriever is first employed to compute the similarity between the documents and the user input. The top-k document set with the highest similarity is then selected as the retrieval result. Finally, both the aforementioned retrieval result and user input are concatenated and fed into the LLMs to generate the response. This approach has achieved significant performance improvements in knowledge-intensive tasks. However, naive RAG encounters several notable issues, including influence of retriever’s capacity, inconsistencies between the LLM’s response and the retrieved documents, and over-reliance on the retrieval result [27]. Numerous efforts have been made to overcome the above downside.

In “retrieval-then-read” framework, the user input is directly used to retrieve relevant documents. However, due to the complexity and ambiguity of language, such as professional terms and ambiguous abbreviations, user input may not always be suitable to serve as the query. Ma et al. [28] proposed the “Rewrite-Retrieve-Read” paradigm, which utilizes LLMs or PLMs to rewrite user inputs, making the rewritten queries more suitable for retrieval tasks, thereby improving the quality of the retrieval results. In practice, redundant information always overwhelms critical information, thereby interfering with the final response of LLM. Moreover, long contexts may lead to the “middle loss” problem [29], where LLMs always focus on the beginning and the end of texts, neglecting the middle sections. Therefore, the retrieved documents need to be further processed to help LLMs better leverage external knowledge. Jiang et al. [30] employed an additional LM to detect and remove unimportant portion. Through compression, the original prompt becomes well understood by LLMs. In the standard RAG framework, the retrieval step is usually performed only once, which only provides limited information [31]. Asai et al. [32] improve the original RAG framework by enabling LLMs to actively determine when and what should be retrieved, thus becoming sufficient for complex multi-step reasoning tasks.

Because of the unique advantage in zero-shot and few-shot learning, LLMs are able to quickly learn and adapt to novel tasks through prompt strategy, although they have never been explicitly trained on these tasks before (Fig. 6). A number of studies have applied LLMs to the medical domain. With the development of society, mental health has attracted increasing attention. Yang et al. [33] utilize LLMs in the mental health analysis task to detect stress, depression, and suicide, achieving impressive performance. Due to the existence of specialized terminology in radiology reports, these reports are often difficult for patients without the medical background to understand. Jeblick et al. [34] employed LLMs to transform the original radiology reports into a version that is easy for patients to understand, allowing them to play a more active role in their own treatment process. Another important application domain is Digital Twins, where LLMs can analyze simulation data from Digital Twins and make fast and intelligent decisions. In smart manufacturing, for example, LLMs detect hazardous signals based on sensor data from Digital Twins, notifying administrator in time. Moreover, LLMs can be employed to perform predictive maintenance and automatically generate fault reports [35], including analysis of fault causes and maintenance history, and predict follow-up maintenance recommendations. In short, combining LLMs with Digital Twins is able to improve the degree of automation.

Interestingly, the 2024 Nobel Prizes in Physics and Chemistry were awarded to AI researchers, which sufficiently show the tremendous potential of AI for science. Because of their impressive capacity, LLMs revolutionized the way how we explore AI for scientific research. To make new discoveries, researchers have to master a preexisting knowledge system that is rapidly growing and increasingly specialized. Such “burden of knowledge” seriously hinders the development of science, since researchers spend too much on reading related literature to keep up with the latest advancements in their fields. Nowadays, researchers turn to LLM-based scientific assistants to help address the above problems. In practice, LLMs are adopted to analyze, extract, and integrate meaningful information from a mount of literature, so as to greatly enhance researchers’ work efficiency. In addition to great analytical and comprehension abilities, LLMs excel at complex reasoning tasks, allowing to accelerate the experimental process. For instance, in drug development, protein structures exhibit regularity, enabling researchers to decompose the prediction process into multiple reasoning steps and design optimal COT strategies for LLMs, thereby speeding up drug molecule discovery [36,37]. By ingesting larger volumes of collected data, LLMs can learn more powerful patterns and regularities within the data, resulting in accurately modeling these complex systems. In environmental and climate science, LLMs can be used for climate modeling and prediction, facilitating researchers to better understand the complex mechanisms of climate change and conduct weather forecasts.

With the continuous exploration of LLMs, their properties are also constantly being revealed [38,39]. Based on these properties, we can conduct a more in-depth analysis of the emergence abilities of LLMs. For example, some works showcase that different parameters in LLMs contribute differently to outcomes [38]. Thus, we can analyze the importance of different parameters during the training stage and investigate the dynamic changes of super weight. The aforementioned experiments allow us to explore the contribution of parameter change to the emergent abilities of LLMs. Moreover, because the knowledge in LLMs is stored in feed-forward network (FFN) modules of each layer [39], the probe can be adopted to monitor FFN modules across different layers during the training phase. Therefore, we can analyze the relationship between the knowledge contained in the LLM and the emergent abilities of LLMs in the dimension of training time. In short, taking the training phase into consideration can provide interesting perspective for figuring out the source of emergent abilities.

For RLHF, the feedback data for human or AI are crucial. However, in some complex scenarios, even humans struggle to effectively judge whether the outputs of LLMs align with human preferences, let alone LLMs themselves. Hence, ensuring that LLMs align with human preferences in complex scenarios is a challenge that remained to be addressed. Given the difficulty of constructing positive evaluation samples, we should shift our focus toward exploring negative samples. Current research tends to treat positive and negative evaluation samples equally. However, negative evaluation samples are generally easier to obtain compared to positive ones. Therefore, using methods such as unlearning [40] or contrastive learning [41], we can guide LLMs to avoid generating negative evaluation outputs, thereby achieving human preference alignment in complex scenarios. Discrimination of human alignment involves multiple aspects. For example, the responses generated by LLMs may conform to human values in one dimension (e.g., provide useful information) but have problems in another dimension (e.g., may introduce bias or be harmful). The reward signals from human annotators or LLMs are often a single scalar, failing to fully represent human preferences. This limitation might lead LLMs to be optimized for wrong objectives, potentially resulting in misaligned outputs. Sophisticated signals or training strategies should be designed to capture fine-grained human alignment criteria.

The external knowledge of most existing works is based on authoritative sources such as Wikipedia. In practice, documents are from a wide range of individual users or self-media organizations, and not verified by professionals or institutions. Therefore, these retrieved documents may contain conflicting information, which hinders LLMs from integrating information from different retrieved documents. Explicitly modeling knowledge conflicts between documents is beneficial for enabling LLMs to better understand and analyze the knowledge contained in retrieved documents, thereby improving the accuracy of generated content.

Current retrieval methods, which select documents from a database based on similarity, is able to obtain documents most relevant to the query. However, the above methods just focus on the gain from an individual retrieved document, resulting in retrieving a collection of documents that are highly similar to one another. Thus, retrieval result is a locally optimal set of documents and fails to offer sufficient external knowledge to the LLMs. To address this issue, the retrieval process should consider the integral benefit of retrieved documents. While utilizing similarity as selection criterion, the knowledge contribution of each candidate document to the retrieved documents should also be evaluated so as to obtain the globally optimal set of retrieved documents.

The cross-domain application of LLMs has achieved remarkable performance. However, in some domains with small data scale, without sufficient data for fine-tuning, LLMs always fail to adequately learn new features and patterns, leading to poor performance. The above problems pose a significant challenge for generalization of LLMs. As one of the promising solutions, counterfactual data augmentation scales dataset with “counterfactual” instance. Based on hypothetical scenarios, counterfactual data augmentation alters conditions or features of input, leading to “counterfactual” samples and enriching dataset. In addition to only scaling data size, counterfactual samples introduce a wide variety of counterfactual perturbations and urge LLMs to explore causal relationships between variables within the data, reducing the LLMs’ dependence on spurious correlation that causes poor generalization. Overall, by facilitating LLMs to learn richer, more robust features and patterns, counterfactual data augmentation improves their abilities of generalization and causal reasoning. In the future, counterfactual data augmentation will be one of the key techniques for cross-domain application.

Since LLMs are essentially black box models, they suffer from the lack of explainability and transparency, and tend to be overconfident. Furthermore, since LLMs are trained on large-scale corpora, which consists of both high-quality and low-quality data, LLMs are likely to generate biased or hallucinated responses. Such limitation significantly impacts the application of LLMs in scientific community. To solve these problems, it is particularly important to conduct in-depth research on the uncertainty estimation for LLMs. Uncertainty estimation approach provides users with confidence information about the predicted results, allowing them to judge the reliability of the model output and avoid misleading conclusions. Besides, in simulation experiments, uncertainty estimation is able to indicate which prediction results need further validation or manual intervention, thereby improving the efficiency of human–machine collaboration. For LLMs, uncertainty estimation is a promising approach for improving reliability and trustworthiness, which attracts increasing attention from researchers.

In recent years, substantial advancements have been achieved in the field of LLMs. Among the prominent developments, DeepSeek-R1 and DeepSeek-V2 emerge as the pioneers that investigate innovative methodologies to promote the evolution of LLMs. Through algorithmic innovation, DeepSeek demonstrates extraordinary performance and showcases its potential to extend the boundaries of natural language understanding and problem-solving.

Due to the huge number of parameters, the requirement of substantial memory footprint during inference hinders deployment of LLMs in real-world applications. Since the proposal of the LLM, boosting inference efficiency has attracted considerable attention. However, these methods often suffer from the problem of degradation, failing to achieve the desired balance between performance and efficiency. To this end, DeepSeek-V2 designs an innovative module, named multi-head latent attention (MLA) [6]. Through mapping the key-value (KV) cache into a latent vector, MLA effectively compresses the computational cost and captures long-distance dependencies. Moreover, DeepSeek-V2 also adopts the DeepSeekMoE [42] architecture, which consists of multiple expert subnetworks. The benefits of multiple expert subnetworks can be 3-fold. First, based on different inputs during inference, the model can dynamically select and activate a small subset of these experts, reducing computational cost. Meanwhile, the overall capacity of the model is also preserved. Second, different experts are assigned to capture diverse linguistic patterns, enabling flexible scaling and task-specific fine-tuning. Finally, by analyzing the activation patterns of each expert subnetwork, researchers can trace the logical pathways behind the decision-making process. Both MLA and DeepSeekMoE show the great potential of architectural optimization and sophisticated mechanisms in driving the advancement of LLMs. As algorithms become more important in the research of LLMs, we should increasingly focus on algorithmic refinement, resource economy, and interpretability measures. This shift is expected to establish the foundation for AI systems that are more accessible, efficient, and controllable, contributing to the sustainable development of robust AI.

Another interesting finding is about distillation. In DeepSeek-R1, the authors fine-tuned smaller LLMs using the data generated by DeepSeek-R1. The experimental results show that fine-tuned LLMs exhibit great reasoning abilities, which are comparable to those of large-scale LLMs. Notably, in specific tasks such as mathematical reasoning and programming, the fine-tuned smaller models even surpass many traditional open-source models. Those findings show that distillation is a promising way to improve the reasoning capabilities of smaller LLMs. Developing a DeepSeek-R1-level model from scratch can be prohibitively expensive, potentially costing millions of dollars, which may discourage many researchers with limited budgets. Distillation offers a more cost-effective option. By designing effective distillation strategies, the reasoning patterns and complex reasoning capacities of larger LLMs are transferred to smaller ones, which is more affordable. In addition to improving the reasoning performance of smaller models, distillation also greatly decreases computational costs during deployment and provides a viable path for deploying efficient reasoning models in resource-constrained environments.

Due to the impressive intelligence capacity, including complex reasoning and understanding, LLMs have attracted public attention. However, the essence of LLMs still remains to be answered. Given our limited understanding of LLMs, people often rely on some simplification approaches such as “labeling” to discuss them. Based on labeling, people use beautiful vocabulary to describe LLMs in the hope that they will eventually become “intelligent individuals”. However, the beautiful “metaphors” make people overlook the ethical and legal risks behind LLMs. On the other hand, some critics hold different opinions, regarding LLMs as systems merely performing “data fitting”. Labeling enables people to quickly understand complex systems, but it is also a double-edged sword, potentially obscuring the ethical and legal risks inherent in LLMs. If we consider LLMs as tools, then the creators of LLMs are engaging in commercial activities. The unauthorized use of public data for commercial activities in the training of LLMs has infringed upon the rights of numerous individuals or institutions. However, if LLMs are regarded as the true “individual”, their training process can be considered as “learning”, just like human, and learning on public data does not require additional payment. Moreover, personifying may blur the boundaries of responsibility between model developers and users, complicating the assignment of responsibility. When a vehicle equipped with an intelligent driving system is involved in a car accident, the owner of the vehicle may blame the intelligent driving system’s “wrong decision-making”, while the developer may claim that the system operates based on training data and predefined rules, and that the accident has more to do with the user’s usage or the specific environment. While embracing the potential of technology, we also need to maintain critical thinking to ensure that technological development always centers around serving human well-being.