The last edition of Microsoft Z Justification of Phi-4 The key challenge is assumed in building artificial intelligence systems capable of reasoning. Since the introduction of the rethinking of thoughts in 2022, scientists have believed that advanced reasoning required very large language models with hundreds of billions of parameters. However, the new 14-million Microsoft parameter, the justification of Phi-4, questions this belief. Using the data -oriented approach, instead of relying on pure computing power, the model reaches comparable performance with much larger systems. This breakthrough shows that data -oriented approach can be equally effective in training of reasoning models as in the case of conventional AI training. This opens the possibility that smaller AI models have achieved advanced reasoning, changing the way AI programmers trains reasoning models, passing from “larger is better” to “better data is better.”
Traditional reasoning paradigm
Chain reasoning It has become a standard of solving complex problems in artificial intelligence. This technique directs the language model by reasoning step by step, dividing difficult problems into smaller steps. He imitates human thinking, making the models “Think loud” in natural language before answering.
However, this skill had an important limitation. Researchers consistently found This chain hints only worked well when language models were very large. The ability to reason seemed directly related to the size of the model, and larger models worked better in complex tasks of reasoning. This discovery led to competition in building large reasoning models in which companies focused on transforming their large language models into powerful reasoning engines.
The idea of including the ability to reason in AI models is primarily due to the observation that large language models can make Learning in context. Researchers noticed That when models have been shown examples of solving problems step by step, they learn to follow this pattern for new problems. This led to the belief that larger models trained in the field of huge data naturally develop more advanced reasoning. A strong relationship between the size of the model and the efficiency of reasoning has become wisdom. The teams have invested huge resources in scaling the ability to reason using reinforcement learning, believing that the computing force was the key to advanced reasoning.
Understanding the focus on data
The increase in artificial -oriented intelligence questions the mentality “more is better”. This approach moves the focus from the architecture of the model to careful data engineering used to train AI systems. Instead of treating data as a fixed input data, the focused methodology perceives data as a material that can be improved and optimized to increase AI efficiency.
Andrew Ng, leader in this field, promotes Building systematic engineering practices to improve the quality of data, not just adjusting the code or scaling models. This philosophy recognizes that data quality and treatment often matter than the size of the model. Companies taking this approach show that smaller, well -trained models can exceed larger ones if they are trained in high quality, carefully prepared data sets.
Data -oriented approach asks another question: “How can we improve our data?” Instead of “how can we increase the model?” This means creating better training sets, improving data quality and developing systematic data engineering. In artificial data intelligence, emphasis is placed on understanding, which makes the data effective for specific tasks, and not just gathering them.
This approach showed a great promise in training small, but powerful AI models using small data sets and much less calculations. Phi Microsoft models are a good example of training of small language models using a data -oriented approach. These models are trained with Learning the curriculum What is primarily inspired by how children learn through gradually more difficult examples. Initially, the models are trained in easy examples, which are then gradually replaced by more difficult. Microsoft has built a set of data from textbooks, as explained in their article “Textbooks are all you need. “It helped to surpass Phi-3 models, such as Gemma Google and GPT 3.5 in tasks such as language understanding, general knowledge, mathematical problems at primary school and answering medical questions.
Despite the success of the approach -oriented approach, the reasoning generally remained a feature of large AI models. This is due to the fact that reasoning requires complex patterns and knowledge that large -scale models are more easily captured. However, this belief has recently been questioned by the development of the Phi-4 justification model.
PHI-4 caming strategy
The victim of Phi-4 shows how you can use a focused approach on data for training small reasoning models. The model was built by the supervised tuning of the basic Phi-4 model on carefully selected “taught” hints and reasoning of examples generated using O3-Mini OpenAI. It was focused on quality and specificity rather than the size of the data set. The model is trained using about 1.4 million high -quality hints instead of billions of general ones. Scientists filter examples to cover different levels of difficulty and types of reasoning, ensuring diversity. This careful treatment meant that every sample example was intentional, teaching the model specific reasoning patterns, and not just an increase in data volume.
As part of the supervised tuning, the model is trained with full reasoning of the demonstration covering the total thought process. These chains of reasoning step by step helped the model learn how to build logical arguments and systematically solve problems. To further increase the ability to reason, it is further improved by learning to strengthen at about 6,000 high -quality mathematical problems with verified solutions. This shows that even small amounts of concentrated reinforcement learning can significantly improve reasoning when used for well -chestnized data.
Performance beyond expectations
The results prove that this data -oriented approach works. Identifying Phi-4 exceeds much larger open models, such as Deepseek-R1-Distill-LaMa-70B And it almost matches the full Deepseek-R1, even though it was much smaller. In the AIME 2025 test (qualifier of the American Mathematics Olympics), receiving Phi-4 Beats Deepseek-R1, which has 671 billion parameters.
These profits go beyond mathematics for scientific problem solving, coding, algorithms, planning and spatial tasks. Improvements from careful transfer of data treatment to general reference points, suggesting that this method builds fundamental reasoning skills, not tricks specific to the task.
He assumes Phi-4 challenges that advanced reasoning requires huge calculations. The 14-million parameter model can match the performance of models tens of times larger after training carefully selected data. This performance has important consequences for the implementation of the reasoning of artificial intelligence in which the resources are limited.
Implications for the development of AI
The success of Phi-4 Opporting signals a change in the way you should build AI reasoning models. Instead of focusing mainly on increasing the size of the model, teams can get better results by investing in data quality and treatment. This means that advanced reasoning is more accessible to organizations without huge computing budgets.
The focused method also opens new research paths. Future work can focus on finding better training hints, introducing richer demonstrations of reasoning and understanding that data helps to reason best. These directions can be more productive than just building larger models.
To put it more, it can help in democratization of artificial intelligence. If smaller models trained in the field of selected data can match large models, advanced artificial intelligence becomes available to a larger number of programmers and organizations. It can also accelerate AI's reception and innovations in areas where very large models are not practical.
The future of reasoning models
PHI-4 planting is a new standard for the development of the reasoning model. Future AI systems will probably balance careful data treatment with architectural improvements. This approach confirms that both the quality of the data and the design of the model is important, but the improvement of data can bring faster, more profitable profits.
This also enables specialized reasoning models trained in the field of data specific to the domain. Instead of general purpose giants, teams can build concentrated perfect models in individual fields by targeted data treatment. This will create a more efficient artificial intelligence for specific applications.
As AI progresses, lessons from the justification of Phi-4 will not only affect the training of the reasoning model, but in general the development of artificial intelligence. The success of data treatment.
Lower line
Microsoft changes in justified Phi-4 together the common belief that AI advanced reasoning requires very large models. Instead of relying on a larger size, this model uses a data -oriented approach from high quality and carefully selected training data. PHI-4's offering has only 14 billion parameters, but also performs much larger models in difficult tasks of reasoning. This shows that focusing on better data is more important than just increasing the size of the model.
This new way of training means that advanced AI is more efficient and available to organizations that do not have large computer resources. The success of the phi-4 justification indicates a new direction of AI development. It focuses on improving the quality of data, intelligent training and careful engineering, and not only on increasing models.
This approach can help in the development of artificial intelligence, reduce costs and allow more people and companies to use powerful AI tools. In the future, artificial intelligence will probably grow, combining better models with better data, thanks to which advanced AI useful in many specialized areas.
