Apple Study Finds AI Reasoning Models Are Not as Smart as Thought

Apple’s new study reveals that AI reasoning models, including those from Meta and OpenAI, are not as capable as believed, suffering from ‘complete accuracy collapse’ when handling complex tasks. This disputes the narrative of nearing artificial general intelligence and suggests that reliance on statistical guesswork often leads to inaccuracies. Many in the AI community view this research as a wake-up call for more realistic assessments of AI capabilities.

A recent study by Apple has sent ripples through the AI community, claiming that reasoning models aren’t as capable as once believed. Researchers found these large language models (LLMs), like those from Meta and OpenAI, suffer a ‘complete accuracy collapse’ when faced with complex tasks, challenging the narrative of advancing toward artificial general intelligence (AGI).

Apple’s findings, published on their Machine Learning Research website on June 7, argue that not only do reasoning models fail at generalized reasoning, but their performance deteriorates significantly under pressure. The study reveals an intriguing scaling limit: as problem complexity increases, these models can initially rise to the challenge but ultimately falter, diminishing their reasoning capabilities.

These models, which are supposed to enhance AI by mimicking step-by-step reasoning—what’s known as the “chain-of-thought” method—often rely on statistical patterns instead of true understanding. While they articulate their thought process, this can lead to ‘hallucinations’, resulting in false answers or disinformation. An OpenAI report backs up these claims, noting that reasoning models are particularly prone to such errors, with hallucination rates skyrocketing as models advance.

In fact, when tasked with summarizing personal facts, OpenAI’s advanced models demonstrated a significant jump in hallucination rates: from 16% with their earlier models to as high as 48% in their latest iterations. OpenAI has acknowledged the need for further inquiry into these findings.

The Apple study critiques existing evaluation methods, asserting they focus too heavily on mathematical metrics and ignore the broader complexities of reasoning tests. To explore this, the researchers tasked various AI models—including those from OpenAI and Google’s Gemini—with classic puzzles like the Tower of Hanoi and River Crossing, adjusting complexity along the way.

At lower complexity levels, the more conventional models generally outperformed reasoning models, but as questions grew in difficulty, reasoning models did initially cooperate better. However, both groups collapsed entirely when faced with extremely complex puzzles.

In a twist, even when researchers provided solution algorithms to the models, their problem-solving abilities didn’t improve. The study revealed unexpected behavioral patterns, as some models completed numerous correct answers on tasks like the Tower of Hanoi but struggled with simpler tasks such as the River Crossing.

These outcomes suggest that reasoning models depend more on recognizing patterns than on actual logically derived conclusions. And while meaningful, the researchers acknowledged limitations in their study, noting that their problems only depict a small fraction of potential reasoning tasks.

Yet, Apple’s position in this rat race is shaky—its Siri functionality has reportedly lagged significantly behind competitors. Some critics suspect this study is a mere attempt to downplay rivals’ efforts, with university professor Pedros Domingos remarking that “Apple’s brilliant new AI strategy is to prove it doesn’t exist.”

Nonetheless, others in the AI field see this research as a critical counter to lofty claims about the prospects of current AI technologies potentially achieving superintelligence. AI expert Andriy Burkov hailed Apple’s work for clarifying that LLMs retain the same limitations as traditional neural networks, advocating that researchers should treat these technologies as mathematical functions rather than engaging with them as one would with a sentient entity.

In the grand scheme of things, this Apple study underscores the urgent need for a more grounded perspective on the capabilities and limitations of AI as the world continues to grapple with evolving technologies. Only time will tell what advancements might come next, but it’s clear that the journey has just begun.

The Apple study exposes significant weaknesses in current AI reasoning models, casting doubt on their ability to perform complex reasoning tasks. This comes amid claims of advancing toward artificial general intelligence. Many in the AI community welcome these findings, suggesting it’s a necessary check against the overblown expectations surrounding AI’s mechanical reasoning capabilities. As the race for AI development continues, the conversation is shifting to a more realistic understanding of what these technologies can and cannot achieve.

Original Source: www.livescience.com