Reasoning Models of Artificial Intelligence Think Like Humans Do

For decades, scientists have debated how the mind solves problems that were never explicitly taught. When a child learns to identify a car from a picture, or when a student figures out a new math problem without being shown the exact steps, it often feels as though a hidden, "derived" connection has suddenly appeared in the brain. Some theories suggest these connections emerge spontaneously, bypassing direct instruction. However, a different perspective argues that nothing truly emerges without a history of reinforcement. This view, known as mediational theory, proposes that what looks like a sudden insight is actually the result of a series of small, intermediate steps. These steps—often silent words or mental images—are practiced and strengthened by their own consequences, even if the teacher or researcher never intended to train them specifically. It is a question of whether the mind builds bridges between ideas through invisible, reinforced habits or through some mysterious leap of logic.

This question has recently moved beyond the study of human behavior and into the heart of artificial intelligence. A new perspective from researchers at Claremont Graduate University, the Florida Institute of Technology, and California State University Los Angeles suggests that the most advanced computer programs are now proving the same point that behavior scientists have long argued. By examining how modern large language models solve complex tasks, the authors show that these machines do not jump straight to an answer. Instead, they engage in a step-by-step process that mirrors the very human behavior described by mediational theory. The researchers argue that the "reasoning" seen in these systems is not magic or a sudden emergence of intelligence, but rather the result of training that reinforces each intermediate step of a solution, just as a human learner reinforces their own internal thoughts.

The paper begins by revisiting the idea of derived relations, which are connections between things that a person has not been directly taught to link. In traditional experiments, a child might be taught that a picture of a car goes with the word "car," and later, without further training, they might correctly point to the car when asked to find the "car" among other objects. This is often called stimulus equivalence. The authors challenge the idea that this link is untrained or un-reinforced. They propose that when the child hears the word "car," they silently or aloud repeat the word to themselves. This act of repeating the word is a mediating behavior. Because the child is in the presence of the car, saying the word "car" is reinforced by the match between what they hear and what they say. This reinforcement happens directly, even if the researcher did not explicitly plan to reward the child for saying the word. The connection is not a mystery; it is a direct result of the child's own behavior being strengthened by the environment.

The authors extend this logic to how computers are now being taught to think. Early versions of artificial intelligence were like simple prediction machines. They would look at a sentence and guess the next word based on statistical patterns, much like a person guessing the next word in a familiar phrase. These systems could answer questions, but they did so in a single, rapid shot, without breaking the problem down. They lacked the ability to reason through a complex issue. The shift toward "reasoning models" occurred when engineers realized that forcing the computer to show its work improved its performance. Instead of just giving an answer, the computer was prompted to generate a chain of thought, a sequence of steps leading to the solution. This change in how the computer was trained aligns perfectly with the mediational theory of human behavior.

The training of these reasoning models happens in stages, each refining how the machine learns. First, the model is exposed to vast amounts of text, learning the basic patterns of language. Then, it is fine-tuned with examples of helpful and safe answers provided by human experts. Finally, and most importantly for this argument, the model undergoes reinforcement learning. In this stage, the computer is not just told if its final answer is right or wrong; it is evaluated on the steps it took to get there. When a model solves a problem correctly, the system reinforces the entire chain of steps that led to that answer. However, in standard models that reward only the final answer, the system reinforces the whole chain as a group, meaning a faulty step might be strengthened if the final answer happens to be correct. To address this, specific "process reward" models have been developed that check the accuracy of each step individually. In these advanced systems, if a step is incorrect, its probability is decreased even if the final answer is correct, ensuring the machine learns to take correct intermediate steps rather than arriving at a correct conclusion by luck.

The researchers point out that this is exactly how human problem-solving works. When a person faces a difficult task, they do not usually produce the final answer in one uninterrupted burst. Instead, they break the problem down, talking to themselves or writing down steps. Each step serves as a cue for the next and is reinforced by the progress it makes toward the solution. The authors note that when computer engineers observed their models suddenly starting to check their own work or correct their own errors, they called it an "emergent" behavior. The researchers argue that this is a misnomer. Nothing emerged that was not reinforced. The models learned to check their work because doing so led to correct answers more often, and the system rewarded those chains of behavior. The "emergence" was simply a gap in the observer's knowledge of the training history, not a gap in the mechanism of action.

This convergence of computer science and behavioral science offers a powerful insight. The engineers who built these systems arrived at the same conclusion as behavior scientists: the most effective way to solve a complex problem is to use a sequence of mediating steps that are reinforced along the way. The paper suggests that this is not a coincidence but a fundamental truth about how problems are solved, whether by a human brain or a silicon chip. The "reasoning" we see in these advanced models is simply the result of a training history that strengthens the intermediate behaviors necessary to reach a goal. By understanding this, we can see that the complex, step-by-step thinking we admire in both humans and machines is built on the same basic principles of learning and reinforcement. The mystery of how we think is not found in a hidden leap of logic, but in the quiet, reinforced steps we take to get there.