Artificial intelligence may have impressive inferencing powers, but don’t count on it to have anything close to human reasoning powers anytime soon. The march to so-called artificial general intelligence (AGI), or AI capable of applying reasoning through changing tasks or environments in the same manner as humans, is still a long way off. Large reasoning models (LRMs), while not perfect, do offer a tentative step in that direction.
In other words, don’t count on your meal-prep service robot to react appropriately to a kitchen fire or a pet jumping on the table and slurping up food.
Also: Meta’s new AI lab aims to deliver ‘personal superintelligence for everyone’ – whatever that means
The holy grail of AI has long been to think and reason as humanly as possible — and industry leaders and experts agree that we still have a long way to go before we reach such intelligence. But large language models (LLMs) and their slightly more advanced LRM offspring operate on predictive analytics based on data patterns, not complex human-like reasoning.
Nevertheless, the chatter around AGI and LRMs keeps growing, and it was inevitable that the hype would far outpace the actual available technology.
“We’re currently in the middle of an AI success theatre plague,” said Robert Blumofe, chief technology officer and executive VP at Akamai. “There’s an illusion of progress created by headline-grabbing demos, anecdotal wins, and exaggerated capabilities. In reality, truly intelligent, thinking AI is a long ways away.”
A recent paper written by Apple researchers downplayed LRMs’ readiness. The researchers concluded that LRMs, as they currently stand, aren’t really conducting much reasoning above and beyond the standard LLMs now in widespread use. (My ZDNET colleagues Lester Mapp and Sabrina Ortiz provide excellent overviews of the paper’s findings.)
Also: Apple’s ‘The Illusion of Thinking’ is shocking – but here’s what it missed
LRMs are “derived from LLMs during the post-training phase, as seen in models like DeepSeek-R1,” said Xuedong Huang, chief technology officer at Zoom. “The current generation of LRMs optimizes only for the final answer, not the reasoning process itself, which can lead to flawed or hallucinated intermediate steps.”
LRMs employ step-by-step chains of thought, but “we must recognize that this does not equate to genuine cognition, it merely mimics it,” said Ivana Bartoletti, chief AI governance officer at Wipro. “It’s likely that chain-of-thought techniques will improve, but it’s important to stay grounded in our understanding of their current limitations.”
LRMs and LLMs are prediction engines, “not problem solvers,” Blumofe said. “Their reasoning is done by mimicking patterns, not by algorithmically solving problems. So it looks like logic, but doesn’t behave like logic. The future of reasoning in AI won’t come from LLMs or LRMs accessing better data or spending more time on reasoning. It requires a fundamentally different kind of architecture that doesn’t rely entirely on LLMs, but rather integrates more traditional technology tools with real-time user data and AI.”
Also: 9 programming tasks you shouldn’t hand off to AI – and why
Right now, a better term for AI’s reasoning capabilities may be “jagged intelligence,” said Caiming Xiong, vice president of AI research at Salesforce. “This is where AI systems excel at one task but fail spectacularly at another — particularly within enterprise use cases.”
What are the potential use cases for LRMs? And what’s the benefit of adopting and maintaining these models? For starters, use cases may look more like extensions of current LLMs. They will arise in a number of areas — but it’s complicated. “The next frontier of reasoning models are reasoning tasks that — unlike math or coding — are hard to verify automatically,” said Daniel Hoske, CTO at Cresta.
Currently, available LRMs cover most of the use cases of classic LLMs — such as “creative writing, planning, and coding,” said Petros Efstathopoulos, vice president of research at RSA Conference. “As LRMs continue to be improved and adopted, there will be a ceiling to what models can achieve independently and what the model-collapse boundaries will be. Future systems will better learn how to use and integrate external tools like search engines, physics simulation environments, and coding or security tools.”
Also: 5 tips for building foundation models for AI
Early use cases for enterprise LRMs include contact centers and basic knowledge work. However, these implementations “are rife with subjective problems,” Hoske said. “Examples include troubleshooting technical issues, or planning and executing a multi-step task, given only higher-level goals with imperfect or partial knowledge.” As LRMs evolve, these capabilities may improve, he predicted.
Typically, “LRMs excel at tasks that are easily verifiable but difficult for humans to generate — areas like coding, complex QA, formal planning, and step-based problem solving,” said Huang. “These are precisely the domains where structured reasoning, even if synthetic, can outperform intuition or brute-force token prediction.”
Efstathopoulos reported seeing solid uses of AI in medical research, science, and data analysis. “LRM research results are encouraging, with models already capable of one-shot problem solving, tackling complex reasoning puzzles, planning, and refining responses mid-generation.” But it’s still early in the game for LRMs, which may or may not be the best path to fully reasoning AI.
Also: How AI agents can generate $450 billion by 2028 – and what stands in the way
Trust in the results coming out of LRMs also can be problematic, as it has been for classic LLMs. “What matters is if, beyond capabilities alone, these systems can reason consistently and reliably enough to be trusted beyond low-stakes tasks and into critical business decision-making,” Salesforce’s Xiong said. “Today’s LLMs, including those designed for reasoning, still fall short.”
This doesn’t mean language models are useless, Xiong emphasized. “We’re successfully deploying them for coding assistance, content generation, and customer service automation where their current capabilities provide genuine value.”
Human reasoning is not without immense flaws and bias, either. “We don’t need AI to think like us — we need it to think with us,” said Zoom’s Huang. “Human-style cognition brings cognitive biases and inefficiencies we may not want in machines. The goal is utility, not imitation. An LRM that can reason differently, more rigorously, or even just more transparently than humans might be more helpful in many real-world applications.”
Also: People don’t trust AI but they’re increasingly using it anyway
The goal of LRMs, and ultimately AGI, is to “build toward AI that’s transparent about its limitations, reliable within defined capabilities, and designed to complement human intelligence rather than replace it,” Xiong said. Human oversight is essential, as is “recognition that human judgment, contextual understanding, and ethical reasoning remain irreplaceable,” he added.
Want more stories about AI? Sign up for Innovation, our weekly newsletter.