Artificial intelligence (AI) reasoning models aren't quite as capable as they appear. In reality, their performance breaks down completely when tasks become too complex, according to researchers at Apple.

Reasoning models like Anthropic's Claude, OpenAI's o3, and DeepSeek's R1 are advanced large language models (LLMs) designed to spend more time and computational resources to deliver more accurate responses compared to standard models.

The emergence of these models has led some major tech companies to make new claims that they may be close to achieving artificial general intelligence (AGI), which refers to systems that can surpass humans in most cognitive tasks.

However, a recent paper published on June 7 on Apple's Machine Learning Research website challenges these assertions and delivers a strong rebuttal to competing firms. According to the study, not only do reasoning models fail to demonstrate generalized reasoning ability, but their reasoning capabilities degrade significantly once tasks reach a certain level of complexity.

"Through extensive testing across various puzzles, we show that leading LRMs experience a total accuracy collapse beyond specific complexity thresholds," the researchers noted. "Additionally, they display an unexpected scaling limitation: their reasoning effort increases with problem difficulty up to a point, then diminishes even when sufficient token capacity is available."

LLMs improve by learning from massive amounts of human-generated data. This allows them to generate probabilistic patterns through their neural networks when prompted.

Related: AI frequently 'hallucinates,' but there's a fix

Sign up for the Live Science daily newsletter now. Reasoning models aim to enhance AI precision using a method known as "chain-of-thought." This involves generating multi-step responses that simulate how humans apply logic to solve problems.

This process enables chatbots to review and refine their reasoning, allowing them to handle more challenging tasks with greater accuracy. During chain-of-thought processing, models articulate their logic step-by-step in natural language, making it easier to trace their decision-making process.

Nevertheless, since this approach relies on statistical inference rather than genuine comprehension, chatbots often produce incorrect answers, make things up when they lack information, and sometimes offer strange or even dangerous advice.

An OpenAI technical report revealed that reasoning models are particularly susceptible to hallucinations—more so than regular models—and the issue worsens as models evolve.

For example, when asked to summarize factual information about individuals, the company’s o3 and o4-mini models generated false content 33% and 48% of the time, respectively, compared to just 16% for the earlier o1 model. OpenAI officials admitted they’re unsure why this occurs, stating that "more research is needed to understand the cause of these results."

"We believe the absence of thorough investigations into these issues stems from shortcomings in current evaluation methods," the authors of Apple's new study wrote. "Most existing evaluations center around well-known math and coding benchmarks, which, although useful, often face data contamination and lack controlled experimental conditions across varying complexities. Furthermore, they don’t provide insights into the structure and quality of the reasoning paths generated."

Peeking inside the black box

To better understand these limitations, the researchers tested both generic and reasoning models—including OpenAI's o1 and o3, DeepSeek R1, Anthropic's Claude 3.7 Sonnet, and Google's Gemini—by assigning them four classic puzzles to solve (river crossing, checker jumping, block-stacking, and The Tower of Hanoi). They could vary the puzzle complexity by increasing the number of elements involved.

In low-complexity scenarios, generic models outperformed reasoning models by solving problems without the added computational burden of reasoning chains. As the puzzles grew more complex, reasoning models initially gained an edge—but this advantage vanished entirely under high-complexity conditions, where both types of models saw their performance drop to zero.

Once a critical threshold was crossed, reasoning models allocated fewer tokens (the basic units models use to break down data) to more complex tasks, indicating that they engaged in less reasoning and faced fundamental limits in maintaining long chains of thought. These limitations persisted even when correct solutions were provided.

"When we gave the solution algorithm for the Tower of Hanoi to the models, their performance on this puzzle did not improve," the authors stated. "Moreover, examining the first incorrect move made by the models revealed surprising behavior. For instance, they could successfully complete up to 100 correct moves in the Tower of Hanoi but struggle to complete more than 5 correct moves in the River Crossing puzzle."

The above is the detailed content of Cutting-edge AI models from OpenAI and DeepSeek undergo 'complete collapse' when problems get too difficult, study reveals. For more information, please follow other related articles on the PHP Chinese website!