The difference between reasoning models and normal models is reasoning models are two steps, to oversimplify it a little they prompt "how would you go about responding to this" then prompt "write the response"

It's still predicting the most likely thing to come next, but the difference is that it gives the chance for the model to write the most likely instructions to follow for the task, then the most likely result of following the instructions - both of which are much more conformant to patterns than a single jump from prompt to response.

This has been known for years, this is the default assumption of how these models work.

You would have to prove that some kind of actual reasoning capacity has arisen as... some kind of emergent complexity phenomenon.... not the other way around.

Corpos have just marketed/gaslit us/themselves so hard that they apparently forgot this.

There are ads but they're subtle enough that you don't recognize them as such.

I can envision a system where an LLM becomes one part of a reasoning AI, acting as a kind of fuzzy "dataset" that a proper neural network incorporates and reasons with, and the LLM could be kept real-time updated (sort of) with MCP servers that incorporate anything new it learns.

But I don't think we're anywhere near there yet.

This sort of thing has been published a lot for awhile now, but why is it assumed that this isn't what human reasoning consists of? Isn't all our reasoning ultimately a form of pattern memorization? I sure feel like it is. So to me all these studies that prove they're "just" memorizing patterns don't prove anything other than that, unless coupled with research on the human brain to prove we do something different.

Because it's a fear-mongering angle that still sells. AI has been a vehicle for scifi for so long that trying to convince Boomers that of won't kill us all is the hard part.

I'm a moderate user for code and skeptic of LLM abilities, but 5 years from now when we are leveraging ML models for groundbreaking science and haven't been nuked by SkyNet, all of this will look quaint and silly.

"It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." -Pamela McCorduck´.
It's called the AI Effect.

As Larry Tesler puts it, "AI is whatever hasn't been done yet.".

You know, despite not really believing LLM "intelligence" works anywhere like real intelligence, I kind of thought maybe being good at recognizing patterns was a way to emulate it to a point...

But that study seems to prove they're still not even good at that. At first I was wondering how hard the puzzles must have been, and then there's a bit about LLM finishing 100 move towers of Hanoï (on which they were trained) and failing 4 move river crossings. Logically, those problems are very similar... Also, failing to apply a step-by-step solution they were given.

So they have worked out that LLMs do what they were programmed to do in the way that they were programmed? Shocking.

But it still manages to fuck it up.

I've been experimenting with using Claude's Sonnet model in Copilot in agent mode for my job, and one of the things that's become abundantly clear is that it has certain types of behavior that are heavily represented in the model, so it assumes you want that behavior even if you explicitly tell it you don't.

Say you're working in a yarn workspaces project, and you instruct Copilot to build and test a new dashboard using an instruction file. You'll need to include explicit and repeated reminders all throughout the file to use yarn, not NPM, because even though yarn is very popular today, there are so many older examples of using NPM in its model that it's just going to assume that's what you actually want - thereby fucking up your codebase.

I've also had lots of cases where I tell it I don't want it to edit any code, just to analyze and explain something that's there and how to update it... and then I have to stop it from editing code anyway, because halfway through it forgot that I didn't want edits, just explanations.

You've hit the nail on the head.

Personally, I wish that there's more progress in our understanding of human intelligence.

Agreed. We don't seem to have a very cohesive idea of what human consciousness is or how it works.

...... So you're saying there's a chance?

Employers who are foaming at the mouth at the thought of replacing their workers with cheap AI:

🫢

This paper doesn’t prove that LLMs aren’t good at pattern recognition, it demonstrates the limits of what pattern recognition alone can achieve, especially for compositional, symbolic reasoning.

Unlike Markov models, modern LLMs use transformers that attend to full contexts, enabling them to simulate structured, multi-step reasoning (albeit imperfectly). While they don’t initiate reasoning like humans, they can generate and refine internal chains of thought when prompted, and emerging frameworks (like ReAct or Toolformer) allow them to update working memory via external tools. Reasoning is limited, but not physically impossible, it’s evolving beyond simple pattern-matching toward more dynamic and compositional processing.

Define reason.

Like humans? Of course not. They lack intent, awareness, and grounded meaning. They don’t “understand” problems, they generate token sequences.

Like what?

I don’t think there’s any search engine better than Perplexity. And for scientific research Consensus is miles ahead.

Reasoning is limited

Most people wouldn't call zero of something 'limited'.

The paper doesn’t say LLMs can’t reason, it shows that their reasoning abilities are limited and collapse under increasing complexity or novel structure.