The Birth of an AI with Both Liberal Arts and Science Brains! The Truth Behind the Super-Smart AI That Switches Its 'Way of Thinking' Based on the Problem

"Why does this AI ponder when asked a philosophy question, but answer 'What is 1234 x 5678?' in 0.1 seconds?" An experimental result that even had an MIT-educated genius programmer scratching their head is now the talk of the town.

In July 2025, Carnegie Mellon University announced a new type of AI called 'Agentic-R1,' which possesses a dual personality, as if it has both a liberal arts and a science brain. It carefully thinks through the interpretation of difficult poetry in words, while instantly processing complex calculations using tools.

Moreover, what is surprising is that it is 1/12th the size of ChatGPT and could potentially be used for the price of a cup of coffee per month.

This technology, which Google and Microsoft have already begun to pay attention to, may completely divide your work into 'thinking parts' and 'processing parts,' dramatically improving productivity. Why has AI suddenly become so 'efficient'?

Easy-to-understand audio delivery on Spotify: 'Rami's AI Tech Radio'

A survey by the World Economic Forum ranked 'curiosity and lifelong learning' as the 5th most important skill for 2030.Curiosity is not a hobby, but an economic necessity. I explain this in detail in my book, 'The Ultimate Skill for the AI Era: Curiosity Power.'

Preface

The evolution of generative AI has gone beyond mere text generation and image creation, reaching a stage where it is expected to act as an 'intellectual partner' that solves complex problems.

However, there is still a fundamental dilemma in its problem-solving ability. Some AIs are good at reading long texts and constructing complex logic, but they can make mistakes even in simple calculations. Other AIs can execute programs and perform precise calculations, but they stop thinking the moment they encounter even a slightly unexpected situation.

In this article, we will address this dilemma head-on and explain the Carnegie Mellon University paper, 'Agentic-R1: Distilled Dual-Strategy Reasoning,' which proposes a groundbreaking approach for AI to solve problems more intelligently and flexibly.

The Dilemma of 'Thinking' and 'Execution' in Modern AI

While today's generative AI has made remarkable progress, its capabilities are not universal. Especially in complex problem-solving scenarios, its approaches are broadly divided into two types, each with its own unique strengths and weaknesses.

The Power and Limitations of Long Chain-of-Thought

Recent high-performance language models have dramatically improved their reasoning capabilities through a technique called 'long-CoT.' This is a method where the AI describes its own thought process step-by-step in natural language, similar to how a human writes down their thought process on paper when solving a difficult problem, repeatedly self-verifying and correcting until it reaches the final answer.

This approach has achieved performance that surpasses conventional models, especially in tasks requiring multi-step reasoning, such as mathematical proof problems and complex logic puzzles.

However, this 'thinker' type approach has weaknesses that cannot be ignored. First, because it relies entirely on natural language for all reasoning, the computational load is extremely high, requiring significant time and computing resources to derive a single answer.

Second, because the thought process is a sequence of words, it is structurally prone to errors, as it struggles with strict numerical calculations and fact-checking. It has also been pointed out that it can fall into 'overthinking,' where it unnecessarily complicates what should be a simple problem, leading to redundant and incorrect conclusions.

The dilemma of the 'thinker' is that while it is good at strategic thinking—grasping the structure of a problem globally and setting a path to a solution—it lacks precision in the execution phase.

Reliability and Rigidity of Tool-Augmented Agents

To compensate for the weaknesses of the 'thinker,' AI known as 'Tool-Augmented Agents' has emerged. This is an approach where the language model does not solve problems alone, but appropriately utilizes external 'tools' such as calculators, search engines, and databases.

In particular, for tasks requiring precise calculations, the AI can automatically generate program code to solve the problem and execute it in a code execution environment (interpreter), which can dramatically improve the accuracy and efficiency of the calculations.

This 'executor' type demonstrates immense power, especially in problems centered on large-scale numerical calculations or algorithmic processing. However, they also have their own limitations.

Because their thinking tends to be constrained by the framework of using tools, it is reported that their performance often drops in abstract or conceptually complex reasoning problems. They are left at a loss in situations where it is not self-evident which tool to use or how to use it to solve the problem.

While the reliability of 'execution' in accurately performing tasks is high, the rigidity of being unable to think flexibly—where tool-dependent thinking becomes a hindrance when faced with unknown problems or abstract questions—can be said to be the challenge for 'executors'.

Towards the Integration of Thinking and Execution

The strategic thinking ability of a 'thinker' and the accurate task execution ability of an 'executor.' Is it not possible to create an AI that possesses both? The paper 'Agentic-R1: Distilled Dual-Strategy Reasoning' published by a research team at Carnegie Mellon University attempts to answer exactly this question.

The goal of this research is to integrate the strengths of an approach specialized in thinking and an approach specialized in tool use into a single model. And, proposed to achieve that goal is,

a new fine-tuning framework called DualDistill. The AI model trained using this framework was named Agentic-R1, the protagonist of this article.

The greatest feature of Agentic-R1 is that it learns the ability to dynamically, or autonomously, select and switch between the two strategies depending on the nature of the problem. For example, when faced with a calculation problem where the answer is expected to be a very large number, it calls upon the 'executor' strategy of code execution itself.

On the other hand, for philosophical questions or problems requiring conceptual understanding, it selects the 'thinker' strategy of thinking carefully on a text basis.

The core value of this research does not lie simply in 'adding' two abilities to an AI. It lies in the attempt to grant the AI a higher-level ability—namely, 'situational judgment ability'—regarding when and how to use those two abilities.

This can be said to be an extremely important step in evolving AI from a mere tool that moves as instructed by humans into an autonomous problem-solving partner that seeks out the optimal solution itself.

While AI research to date has progressed along separate axes of 'improving thinking ability' or 'improving tool-use ability,' this research has set 'strategic decision-making' itself, which oversees those abilities, as the target of learning.

This is an attempt to rethink AI intelligence at a higher dimension, and beyond that, a path toward Artificial General Intelligence (AGI) may even come into view.

Technology to Teach 'Situational Judgment' to AI

So, how do we teach 'situational judgment ability' to an AI? The secret lies in the creative learning framework called DualDistill proposed in this research. To understand this method, let us first look at the technology that forms its foundation: 'knowledge distillation'.

What is Knowledge Distillation?

Knowledge distillation is a technology for efficiently transferring the knowledge held by a very large, high-performance AI model (called a 'teacher model') to a lighter, more compact AI model (a 'student model').

Through this process, the student model can demonstrate high performance comparable to the teacher model, despite its smaller size.

What is unique about knowledge distillation is that it does not just teach the student the correct answer (label) to a problem. It also has the student model learn richer information, such as the thought process of 'why the teacher model arrived at that answer' and the 'confidence' (technically called probability distribution) regarding the final answer.

This makes it possible to transfer not only the correct answer itself but also implicit knowledge such as the teacher model's 'way of thinking' and 'criteria for judgment.' DualDistill is a method that applies this concept of knowledge distillation in an unprecedented and creative way.

Learning from Expert Knowledge

In DualDistill, to train a student model, two 'teacher models' with different capabilities are intentionally prepared.

By learning from these two contrasting experts simultaneously, the student model aims to absorb the strengths of both in a balanced manner and integrate them into its own capabilities.

A 'Hybrid Textbook' That Learns from Success and Failure

The most innovative and clever part of DualDistill is this process called 'trajectory synthesis.' This is the step of skillfully combining the solution processes of the two teachers to create a 'hybrid textbook' for the student model. The procedure is as follows.

The true value of this 'trajectory synthesis' lies in the fact that it does not simply teach the correct answer, but rather explicitly structures the 'recovery process from failure' within the learning data. Case 1, in particular, provides the student model with an extremely important meta-level lesson: 'When you make a mistake like this, you should switch to this strategy.'

This becomes a highly effective learning signal for instilling in the AI the self-correction ability to reach the correct answer after trial and error, and the strategic thinking to explore other paths when hitting a dead end. Just as humans learn much from failure, we let the AI systematically learn from failure experiences. This is the breakthrough brought about by 'trajectory synthesis.'

Self-Distillation: Reflecting on and Improving One's Own Capabilities

After learning hybrid strategies from the two teachers, the student model further refines its capabilities through a process called 'self-distillation.' This is, so to speak, a stage where the student studies on its own, views its own abilities objectively, and makes improvements.

Specifically, the student model first attempts the problem set used for learning again and generates multiple solution trajectories. Then, it evaluates the accuracy of those solutions. If it is determined that the student model cannot always answer a problem correctly (i.e., it has not yet fully mastered it), that solution trajectory becomes the target of self-distillation.

This self-distillation process is essential for resolving the 'capability mismatch' that exists between the teacher and the student. For example, even though the student model's tool-use skills are not yet as high as the teacher's, it may try to imitate the teacher and use tools too easily, leading to failure instead.

Self-distillation helps the student model learn strategy selection that matches its current ability, such as making the judgment, 'My tool-use ability is still low, so I will solve this problem with reliable text reasoning.'

This two-stage learning process (teacher distillation → self-distillation) can be interpreted as embodying the concept of 'scaffolding' in educational theory. If the initial 'trajectory synthesis' is hands-on guidance (scaffolding) by an expert teacher, 'self-distillation' is the stage of independence where the scaffolding is gradually removed, and the student solves problems on their own and learns from their own mistakes.

Through this process, knowledge is truly internalized, and the student model grows not into a mere imitator of the teacher, but into an autonomous problem solver.

The 'Smart Strategy Selection' Demonstrated by Experiments

Through the unique training method called DualDistill, has Agentic-R1 truly acquired the 'ability to judge situations'? In this chapter, we will take a detailed look at the results of experiments that verified its capabilities.

Experimental Setup: Benchmarks That Test True Worth

To test the true worth of the model, the research team intentionally prepared challenging datasets for evaluation that cannot be solved by simple thinking or calculation skills alone.

The selection of these benchmarks itself is carefully designed to verify the research hypothesis that a hybrid strategy that flexibly combines 'thinking' and 'execution' is effective, especially for difficult problems.

Results: A Landslide Victory for the Hybrid Strategy

The experimental results clearly showed that Agentic-R1 achieved the performance improvements intended by the research team.

In DeepMath-L and Combinatorics300, where tool use is particularly important, the research model Agentic-R1-SD (the final version that completed self-distillation) recorded scores that significantly exceeded models specialized for tool use and models specialized for pure reasoning.

Furthermore, it is noteworthy that it also surpassed other models and achieved the highest performance in MATH500, a standard benchmark. This suggests that Agentic-R1 has acquired the ability to be an 'excellent generalist' capable of handling a wide range of problems, rather than an 'expert' strong only in specific types of problems.

Also, as evidence that Agentic-R1 is actually performing 'situational judgment,' it has been reported that the tool usage rate changes depending on the nature of the problem.

For example, while it used tools to solve 79.2% of problems in Combinatorics300, which involves many complex calculations, the tool usage rate dropped to 52.0% for the relatively simple AMC dataset. This confirms that the model is not using tools blindly, but is judging their use according to necessity.

Why is 'Trajectory Synthesis' Important?

Is the success of DualDistill really thanks to the clever method called 'trajectory synthesis'? Or did performance simply improve because it learned diverse data from two types of teachers?

To answer this question, the research team conducted a comparative experiment called an 'ablation study.' This is an experiment to investigate how much performance changes when a certain element of the method (in this case, 'trajectory synthesis') is removed.

Specifically, they compared the performance of Agentic-R1 with a model that did not perform 'trajectory synthesis' but simply mixed the solution trajectories generated by the two teachers for learning. The result was clear. Agentic-R1 using 'trajectory synthesis' showed significantly improved performance across all benchmarks.

This result scientifically proves that the success of DualDistill is not just because it was exposed to diverse data, but stems from the method of 'trajectory synthesis' itself, which structured and taught the 'context' of strategy switching.

This can be said to be extremely strong evidence supporting the validity of the most important technical contribution of this paper. The structured lesson of 'This problem cannot be solved with strategy A, so switch to strategy B' was the key to dramatically increasing the model's general problem-solving ability.

AI That Can Learn Even from Mediocre Teachers

Another interesting and thought-provoking finding in this research is that teachers do not necessarily need to be perfect. The agent teacher that took on the role of the 'executor' in the experiment was, in fact, a 'mediocre teacher' in some tasks, with lower performance than the base model used for the student model.

Nevertheless, the student model, Agentic-R1, effectively learned the beneficial strategy of tool use from that mediocre teacher and ultimately achieved results that surpassed the teacher's performance. This suggests a very optimistic and practical possibility in AI development.

In other words, there is no need to wait for the development of a single all-purpose teacher model that can handle every task perfectly; instead, by distilling knowledge from numerous teacher models that are specialized in specific domains and possess 'imperfect but useful skills,' more powerful and versatile AI can be efficiently cultivated.

This is an economically rational approach that promotes modularization and decentralization in AI development.

For example, it opens the door to utilizing various specialized (yet overall imperfect) AIs as a 'teaching staff,' such as an 'AI that is good at coding but bad at writing' or an 'AI that has rich legal knowledge but is poor at calculations.'

This may open up a future where more diverse and robust intelligence can be cultivated more efficiently across the entire AI development ecosystem.

Discussion

What position does Agentic-R1 and its learning method, DualDistill, occupy in the vast map of AI research? In this chapter, we will clarify its novelty through comparison with previous research and discuss the remaining challenges and the future that this research opens up.

Comparison with Previous Research: What is New?

The idea of having an AI use tools itself is not new. For example, previous research such as Toolformer proposes a method where language models learn how to use APIs (gateways for calling external tools) through self-supervised learning (learning without human labeling).

However, these studies mainly focused on when to call an API within a relatively short chain of thought.

On the other hand, long-CoT models, represented by DeepSeek-R1, have been developed with a primary focus on generating the long reasoning process itself and improving its quality, rather than on tool use.

The true novelty of DualDistill lies in the fact that it integrates these two streams. Specifically, the following two points can be mentioned.

This is an approach that draws a line from mere learning of tool usage methods or the deepening of reasoning capabilities. It presents a new paradigm that makes AI think about the strategy itself—not just 'what to do,' but 'how to solve the problem.'

Points of Discussion and Future Challenges

Although Agentic-R1 achieved great success, the paper frankly acknowledges that this approach is not the final form. Several challenges remain for future research.

Furthermore, another interesting point of discussion emerges from the analysis in this paper. It is the trade-off between versatility and specialization. Before performing self-distillation, Agentic-R1 was inferior in performance to pure reasoning models on relatively simple benchmarks (MATH500).

This suggests that the hybrid ability to switch between multiple strategies may create an overhead (extra burden) in judging 'which strategy to use' for simple problems, potentially hindering performance instead.

Elucidating the detailed mechanism of how self-distillation solved this problem and improved performance will be an important research topic in the future.

The Road to General-Purpose Agents

These challenges indicate that this research is not the final goal of AI development, but an important milestone toward the realization of more sophisticated general-purpose agents. The path shown by Agentic-R1 and DualDistill is a concrete and promising framework for AI to acquire more human-like problem-solving capabilities.

When we humans face complex problems, we naturally switch between time spent thinking intuitively and time spent relying on tools like calculators or computers to perform accurate calculations.

This research presented a compelling methodology for teaching AI that kind of 'metacognitive' behavior—that is, the ability to view one's own thinking objectively and select the optimal strategy.

In the future, by further developing this approach, we can expect the realization of autonomous AI agents that powerfully support humans in various fields where diverse thinking and accurate execution are required simultaneously.

For example, we can see a future beyond this research where AI acts as our intellectual partner, such as a researcher formulating and verifying new scientific hypotheses, a software developer designing and debugging complex systems, or a manager driving projects while responding to unexpected problems.

Afterword

In this article, we introduced the challenging research 'DualDistill,' which enables AI models to choose the optimal solution style themselves, and the resulting Agentic-R1 model. The ability to utilize tools and the ability to think through problems logically may seem contradictory at first glance, but both are essential elements in human problem-solving.

The attempt to give AI this duality could be called a step toward Artificial General Intelligence (AGI). Of course, at this point, it is limited to verification in the mathematical domain, but in the future, it is expected that AI that can spontaneously choose problem-solving strategies according to the situation will appear in various scenes.

Readers, too, likely judge whether to think carefully with your own head or rely on tools like calculators or the internet when you face difficult problems in your daily life.

Future AI may also possess such human-like judgment and become our good partner. I hope this article helps you think about a part of cutting-edge AI research and the impact it could have on society.

I hope the ideas and knowledge gained through this article will be of some help to your business. If you felt this article was helpful, I would be encouraged if you could 'like' or 'follow' it. I will continue to share practical know-how and the latest AI trends, so I would be happy if you continue to read.

⚠️ Important: Less than 30% of people can master ChatGPT
Why? It's the difference in 'questioning ability'.
Actually, there is a scientific law to 'how to ask questions' that extracts 10 times the value from AI. A method proven by Google's 20% rule and recommended by Amazon's Bezos. All revealed in 'The Strongest Skill in the AI Era: Curiosity Power'.
Now that AI returns 'answers' instantly, the turning point is in the 'question'. Please check out my book 'The Strongest Skill in the AI Era: Curiosity Power', which condenses the secrets and 13 practical tools you can use in your work tomorrow.

A must-read book for modern people to understand the 'why' of the digital world
Princeton University's legendary professor Brian Kernighan carefully explains the basics of computer science while avoiding technical jargon. Even without programming experience, you can understand the core of technology, from the internet, hardware, and software to AI and security.
The second edition covers the latest topics and is structured to acquire the digital literacy essential in modern society. It is a book packed with knowledge for business people and students alike to correctly understand and wisely utilize the mechanisms of the digital world.

The definitive guide that serves as a compass for the generative AI era. The Matsuo-Iwasawa Laboratory at the University of Tokyo has gathered its full strength to thoroughly explain the latest trends in generative AI. Leading Japanese AI researchers comprehensively analyze the current state and future of generative AI technology, which has evolved rapidly since the advent of ChatGPT.

Published in March 2025, it comprehensively explains information you need to know now, from the core of technology to business applications, domestic and international laws and regulations, and safety. It also covers the latest trends such as AI agents and multimodal capabilities.
A must-have book for everyone involved in the generative AI business. It is a reliable source of information for making sound decisions in the turbulent AI industry.

Why do highly educated people fall for scams and experts make elementary mistakes? This book scientifically elucidates the shocking truth of 'the traps of thinking that the more intelligent you are, the more likely you are to fall into'.
Author Robson analyzes a wealth of cases, from the failures of Nobel laureates to the misdiagnoses of doctors. He exposes the true nature of 'cognitive bias' that cannot be prevented by IQ or academic background alone. The paradox that smart people are more likely to make big mistakes by overestimating their knowledge and neglecting critical thinking.
What this book presents is not just a warning. It is full of concrete methods to acquire true wisdom, such as 'metacognition' and 'intellectual humility'. It is packed with practical wisdom useful in all situations, from business decision-making and investment judgments to daily life.
A book that should be read by those who think 'I'm fine'. By knowing the limits of intelligence, you will acquire true intelligence.