Decoder only LLMs are Markov chains with sophisticated models of the state space. Anyone familiar with Hamiltonian Markov Chains will know that for good results you need a warm up period so that you're sampling from the typical set which is the area with generally the highest probability density in the distribution (not necessary the high propbability/maximum likelihood).
I have spent a lot of time experimenting with Chain of Thought professionally and I have yet to see any evidence to suggest that what's happening with CoT is any more (or less) than this. If you let the model run a bit longer it enters a region close to the typical set and when it's ready to answer you have a high probability of getting a good answer.
There's absolutely no "reasoning" going on here, except that some times sampling from the typical set near the region of your answer is going to look very similar to how human reason before coming up with an answer.
If I'm using an MCMC algorithm to sample a probability distribution, I need to wait for my Markov chain to converge to a stationary distribution before sampling, sure.
But in no way is 'a good answer' a stationary state in the LLM Markov chain. If I continue running next-token prediction, I'm not going to start looping.
I think you're confusing the sampling process and the convergence of those samples with the warmup process (also called 'burn-in') in HMC. When doing HMC MCMC we typically don't start sampling right away (or, more precisely we throw out those samples) because we may be initializing the sampler in a part of the distribution that involves pretty low probability density. After the chain has run awhile it tends to end up sampling from the typical set which, especially in high dimensional distribution, tends to more correctly represent the distribution we actually want to integrate over.
So for language when I say "Bob has three apples, Jane gives him four and Judy takes two how many apples does Bob have" we're actually pretty far from the part of the linguistic manifold where the correct answer is likely to be. As the chain wanders this space it's getting closer until it finally statistically follow the path "this answer is..." and when it's sampling from this path it's in a much more likely neighborhood of the correct answer. That is, after wandering a bit, more and more of the possible paths are closer to where the actual answer lies than they would be if we had just forced the model to choose early.
edit: Michael Betancourt has great introduction to HMC which covers warm-up and the typical set https://arxiv.org/pdf/1701.02434 (he has a ton more content that dives much more deeply into the specifics)
All of this burn-in stuff is designed to get your Markov chain to forget where it started.
But I don’t want to get from “how many apples does Bob have?” to a state where Bob and the apples are forgotten. I want to remember that state, and I probably want to stay close to it — not far away in the “typical set” of all language.
Are you implicitly conditioning the probability distribution or otherwise somehow cutting the manifold down? Then the analogy would be plausible to me, but I don’t understand what conditioning we’re doing and how the LLM respects that.
Or are you claiming that we want to travel to the “closest” high probability region somehow? So we’re not really doing burn-in but something a little more delicate?
You need to think about 1) the latent state 2) the fact that part of the model is post trained to bias the MC towards abiding by the query in the sense of the reward.
A way to look at it is that you effectively have 2 model "heads" inside the LLM, one which generates, one which biases/steers.
The MCMC is initialised based on your prompt, the generator part samples from the language distribution it has learned, while the sharpening/filtering part biases towards stuff that would be likely to have this MCMC give high rewards in the end. So the model regurgitates all the context that is deemed possibly relevant based on traces from the training data (including "tool use", which then injects additional context) and all those tokens shift the latent state into something that is more and more typical of your query.
Importantly, attention acts as a Selector and has multiple heads, and these specialize, so (simplified) one head can maintain focus on your query and "judge" the latent state, while the rest can follow that Markov chain until some subset of the generated+tool injected tokens give enough signal to the "answer now" gate that the middle flips into "summarizing" mode, which then uses the latent state of all of those tokens to actually generate the answer.
So you very much can think of it as sampling repeatedly from an MCMC using a bias, A learned stoping rule and then having a model creating the best possible combination of the traces, except that all this machinery is encoded in the same model weights that get to reuse features between another, for all the benefits and drawbacks that yields.
There was a paper close when OF became a thing that showed that instead of doing CoT, you could just spend that token budget on K parallel shorter queries (by injecting sth. Like "ok, to summarize" and "actually" to force completion ) and pick the best one/majority vote. Since then RLHF has made longer traces more in distribution (although there's another paper that showed as of early 2025 you were trading reduced variance and peak performance as well as loss of edge cases for higher performance on common cases , although this might be ameliorated by now) but that's about the way it broke down 2024-2025
The warmup process is necessary in order to try to find high-probability regions of the target distribution. That's not an issue for an LLM, since it's trained to sample directly from a distribution which looks like natural language.
There is some work on using MCMC to sample from higher-probability regions of an LLM distribution [1], but that's a separate thing. Nobody doubts that an LLM is sampling from its target distribution from the first token it outputs.
> When doing HMC MCMC we typically don't start sampling right away (or, more precisely we throw out those samples) because we may be initializing the sampler in a part of the distribution that involves pretty low probability density.
And how that applies to LLMs? Since they don't do MCMC.
Right, you're describing sampling a single token which is equivalent to sampling from one step in the Markov Chain. When generating output you're repeating this process and updating your state sequentially which is the definition of the Markov Chain since at each state the embedding (which represents our current state) is conditionally independent of the past.
Every response from an LLM is essentially the sampling of a Markov chain.
How does MC warm-up fit with LLMs? With LLMs you start with a prompt, so I don't see how "warm up" applies.
You're not just sampling from them like some MC cases.
> If you let the model run a bit longer it enters a region close to the typical set and when it's ready to answer you have a high probability of getting a good answer.
What does "let the model run a bit longer" even mean in this context?
The way almost all modern LLMs work is by starting with a "thinking" phase which is explicitly not part of the output. You let that process run longer.
This is Dual Process Theory [0] otherwise known as Fast vs Slow thinking, or System 1 and System 2 thinking.
Humans are the only known organism to do System 2 (which doesn't mean we're the only ones that do it, just that we don't know if whales do it), but System 2 is what the author is talking about when they refer to Chains of Thought.
System 1 is what they're referring to when they talk about Messi reacting to an unusual situation on the field.
Related anecdote: I tested myself for ADHD by taking amphetamines. I normally think by intuitive leaps from point to point, without doing the intermediate steps consciously. I found that during this experience my System 2 thinking was fast enough to follow and I actually experienced proper chains of thought. Or I was whizzing my tits off and hallucinated the whole thing. Not sure yet. I should repeat the experiment.
> I tested myself for ADHD by taking amphetamines. [...] Or I was whizzing my tits off and hallucinated the whole thing. Not sure yet.
You can't test yourself for ADHD by taking amphetamines because they have profound effects on everyone - there's a reason why stimulants are some of the most popular recreational drugs. They can make everyone feel smarter, more productive, like you're on top of the world, especially at recreational doses.
When it comes to therapeutic doses for treating ADHD, the effects of the medication should be very subtle, you shouldn't "feel it working" like you've just taken that pill from Limitless. It can produce some euphoria or other recreational effects initially but they're supposed to disappear within a couple of days or weeks, otherwise your dosage is considered to be too high.
> I tested myself for ADHD by taking amphetamines.
You cannot test yourself for ADHD - you need to do the actual work that is required for a proper diagnose. And you cannot test yourself for anything at all by taking amphetamines. Seriously. You were using it recreationally. When you post stuff like this online, people that believe what you write could get hurt.
The irony is I am pretty sure you believe this because of some kind of heuristic, but I am not sure what that heuristic would be called.
Priming has issues with repeatability in experiments.
To believe that "debunks" the entirety of heuristics and bias is completely absurd. The author literally won a Nobel prize for Prospect Theory.
I think that book sucks personally. Just an incredibly dull book to read. The online social contagion that it is a work of fiction might actually be more interesting.
Interesting that he came to this conclusion (CoT should be done in latent space) well before the release of OpenAI's o1, which made explicit CoT reliable in the first place. At the time the blog post was written, CoT was only achieved via a "reason step by step" instruction, which was highly error prone compared to modern o1-like reasoning. (And before InstructGPT/ChatGPT, it was achieved by prompting the model with "let me reason step by step".)
Isn't this a form of intuitive intelligence which doesn't rely on "reasoning"? To me "reasoning" sounds like intentionally trying to solve some explicit problem, while another form of ... insight is the ability to figure out that something is a problem in the first place.
That's, by the way, something LLMs are very much not good at. They possess a superhuman amount of knowledge covering all areas of academia, including math, science, philosophy, engineering, computer science, social sciences and so on, but that doesn't cause them to come up with novel hypotheses and theories. Something that would be easy for a smart human even with a fraction of the academic knowledge of an LLM.
Intentional typos and deliberately misspelling words.. I have thought... is going to have to become a style. Everyone will have their own way of fudging it up.
Sort of like the painters who came after the "Realism" era.
Yann LeCun tis saying that some Paul Halpern from X alleges that Albert Einstein (the one from the stories!) had a conversation with some psychologist more than a hundred years ago where those words were uttered.
If true, this is somehow meant to be taken as a rule/law that language and thoughts are fundamentally different things.
Decoder only LLMs are Markov chains with sophisticated models of the state space. Anyone familiar with Hamiltonian Markov Chains will know that for good results you need a warm up period so that you're sampling from the typical set which is the area with generally the highest probability density in the distribution (not necessary the high propbability/maximum likelihood).
I have spent a lot of time experimenting with Chain of Thought professionally and I have yet to see any evidence to suggest that what's happening with CoT is any more (or less) than this. If you let the model run a bit longer it enters a region close to the typical set and when it's ready to answer you have a high probability of getting a good answer.
There's absolutely no "reasoning" going on here, except that some times sampling from the typical set near the region of your answer is going to look very similar to how human reason before coming up with an answer.
I don't understand the analogy.
If I'm using an MCMC algorithm to sample a probability distribution, I need to wait for my Markov chain to converge to a stationary distribution before sampling, sure.
But in no way is 'a good answer' a stationary state in the LLM Markov chain. If I continue running next-token prediction, I'm not going to start looping.
I think you're confusing the sampling process and the convergence of those samples with the warmup process (also called 'burn-in') in HMC. When doing HMC MCMC we typically don't start sampling right away (or, more precisely we throw out those samples) because we may be initializing the sampler in a part of the distribution that involves pretty low probability density. After the chain has run awhile it tends to end up sampling from the typical set which, especially in high dimensional distribution, tends to more correctly represent the distribution we actually want to integrate over.
So for language when I say "Bob has three apples, Jane gives him four and Judy takes two how many apples does Bob have" we're actually pretty far from the part of the linguistic manifold where the correct answer is likely to be. As the chain wanders this space it's getting closer until it finally statistically follow the path "this answer is..." and when it's sampling from this path it's in a much more likely neighborhood of the correct answer. That is, after wandering a bit, more and more of the possible paths are closer to where the actual answer lies than they would be if we had just forced the model to choose early.
edit: Michael Betancourt has great introduction to HMC which covers warm-up and the typical set https://arxiv.org/pdf/1701.02434 (he has a ton more content that dives much more deeply into the specifics)
No, I still don’t understand the analogy.
All of this burn-in stuff is designed to get your Markov chain to forget where it started.
But I don’t want to get from “how many apples does Bob have?” to a state where Bob and the apples are forgotten. I want to remember that state, and I probably want to stay close to it — not far away in the “typical set” of all language.
Are you implicitly conditioning the probability distribution or otherwise somehow cutting the manifold down? Then the analogy would be plausible to me, but I don’t understand what conditioning we’re doing and how the LLM respects that.
Or are you claiming that we want to travel to the “closest” high probability region somehow? So we’re not really doing burn-in but something a little more delicate?
You need to think about 1) the latent state 2) the fact that part of the model is post trained to bias the MC towards abiding by the query in the sense of the reward.
A way to look at it is that you effectively have 2 model "heads" inside the LLM, one which generates, one which biases/steers.
The MCMC is initialised based on your prompt, the generator part samples from the language distribution it has learned, while the sharpening/filtering part biases towards stuff that would be likely to have this MCMC give high rewards in the end. So the model regurgitates all the context that is deemed possibly relevant based on traces from the training data (including "tool use", which then injects additional context) and all those tokens shift the latent state into something that is more and more typical of your query.
Importantly, attention acts as a Selector and has multiple heads, and these specialize, so (simplified) one head can maintain focus on your query and "judge" the latent state, while the rest can follow that Markov chain until some subset of the generated+tool injected tokens give enough signal to the "answer now" gate that the middle flips into "summarizing" mode, which then uses the latent state of all of those tokens to actually generate the answer.
So you very much can think of it as sampling repeatedly from an MCMC using a bias, A learned stoping rule and then having a model creating the best possible combination of the traces, except that all this machinery is encoded in the same model weights that get to reuse features between another, for all the benefits and drawbacks that yields.
There was a paper close when OF became a thing that showed that instead of doing CoT, you could just spend that token budget on K parallel shorter queries (by injecting sth. Like "ok, to summarize" and "actually" to force completion ) and pick the best one/majority vote. Since then RLHF has made longer traces more in distribution (although there's another paper that showed as of early 2025 you were trading reduced variance and peak performance as well as loss of edge cases for higher performance on common cases , although this might be ameliorated by now) but that's about the way it broke down 2024-2025
The warmup process is necessary in order to try to find high-probability regions of the target distribution. That's not an issue for an LLM, since it's trained to sample directly from a distribution which looks like natural language.
There is some work on using MCMC to sample from higher-probability regions of an LLM distribution [1], but that's a separate thing. Nobody doubts that an LLM is sampling from its target distribution from the first token it outputs.
[1] https://arxiv.org/abs/2510.14901
> When doing HMC MCMC we typically don't start sampling right away (or, more precisely we throw out those samples) because we may be initializing the sampler in a part of the distribution that involves pretty low probability density.
And how that applies to LLMs? Since they don't do MCMC.
In RNNs and Transformers we obtain probability distribution of target variable directly and sample using methods like top-k or temprature sampling.
I don't see the equivalence to MCMC. It's not like we have a complex probability function that we are trying to sample from using a chain.
It's just logistic regression at each step.
Right, you're describing sampling a single token which is equivalent to sampling from one step in the Markov Chain. When generating output you're repeating this process and updating your state sequentially which is the definition of the Markov Chain since at each state the embedding (which represents our current state) is conditionally independent of the past.
Every response from an LLM is essentially the sampling of a Markov chain.
I wish it got called "scaffolding" instead
When you adopt the probability distribution point of view, this is often called 'burn-in'. See eg the usage in https://en.wikipedia.org/wiki/Metropolis%E2%80%93Hastings_al...
That sounds a lot like bad marketing. Chain of thoughts is better, it makes you think the thing is thinking !
How does MC warm-up fit with LLMs? With LLMs you start with a prompt, so I don't see how "warm up" applies.
You're not just sampling from them like some MC cases.
> If you let the model run a bit longer it enters a region close to the typical set and when it's ready to answer you have a high probability of getting a good answer.
What does "let the model run a bit longer" even mean in this context?
The way almost all modern LLMs work is by starting with a "thinking" phase which is explicitly not part of the output. You let that process run longer.
"Obviously no, this IMO proves that we humans can reason efficiently without an inner speech."
Well, no, it proves that Messi can reason efficiently without an inner speech.
This is Dual Process Theory [0] otherwise known as Fast vs Slow thinking, or System 1 and System 2 thinking.
Humans are the only known organism to do System 2 (which doesn't mean we're the only ones that do it, just that we don't know if whales do it), but System 2 is what the author is talking about when they refer to Chains of Thought.
System 1 is what they're referring to when they talk about Messi reacting to an unusual situation on the field.
Related anecdote: I tested myself for ADHD by taking amphetamines. I normally think by intuitive leaps from point to point, without doing the intermediate steps consciously. I found that during this experience my System 2 thinking was fast enough to follow and I actually experienced proper chains of thought. Or I was whizzing my tits off and hallucinated the whole thing. Not sure yet. I should repeat the experiment.
[0] https://en.wikipedia.org/wiki/Dual_process_theory
> I tested myself for ADHD by taking amphetamines. [...] Or I was whizzing my tits off and hallucinated the whole thing. Not sure yet.
You can't test yourself for ADHD by taking amphetamines because they have profound effects on everyone - there's a reason why stimulants are some of the most popular recreational drugs. They can make everyone feel smarter, more productive, like you're on top of the world, especially at recreational doses.
When it comes to therapeutic doses for treating ADHD, the effects of the medication should be very subtle, you shouldn't "feel it working" like you've just taken that pill from Limitless. It can produce some euphoria or other recreational effects initially but they're supposed to disappear within a couple of days or weeks, otherwise your dosage is considered to be too high.
> I tested myself for ADHD by taking amphetamines.
You cannot test yourself for ADHD - you need to do the actual work that is required for a proper diagnose. And you cannot test yourself for anything at all by taking amphetamines. Seriously. You were using it recreationally. When you post stuff like this online, people that believe what you write could get hurt.
Dont know why you are getting the downvotes this is the best comment here.
Haven't many of the fast/slow thinking claims, as popularized by the best-selling Daniel Kahneman's book, been debunked?
The irony is I am pretty sure you believe this because of some kind of heuristic, but I am not sure what that heuristic would be called.
Priming has issues with repeatability in experiments.
To believe that "debunks" the entirety of heuristics and bias is completely absurd. The author literally won a Nobel prize for Prospect Theory.
I think that book sucks personally. Just an incredibly dull book to read. The online social contagion that it is a work of fiction might actually be more interesting.
Sounds like great fodder for a just as informative reply.
Interesting that he came to this conclusion (CoT should be done in latent space) well before the release of OpenAI's o1, which made explicit CoT reliable in the first place. At the time the blog post was written, CoT was only achieved via a "reason step by step" instruction, which was highly error prone compared to modern o1-like reasoning. (And before InstructGPT/ChatGPT, it was achieved by prompting the model with "let me reason step by step".)
Fair to say he demonstrates good CoT reasoning capabilities?
Isn't this a form of intuitive intelligence which doesn't rely on "reasoning"? To me "reasoning" sounds like intentionally trying to solve some explicit problem, while another form of ... insight is the ability to figure out that something is a problem in the first place.
That's, by the way, something LLMs are very much not good at. They possess a superhuman amount of knowledge covering all areas of academia, including math, science, philosophy, engineering, computer science, social sciences and so on, but that doesn't cause them to come up with novel hypotheses and theories. Something that would be easy for a smart human even with a fraction of the academic knowledge of an LLM.
There's a good deal of work on latent reasoning models, these days. Here is a recent survey paper.
https://arxiv.org/abs/2507.06203
> chains of thoughts
Pedantic maybe -- but does this need two plurals?
There's lots of typos like this in the article which I find quite nice for some reason
Reason being the human touch as opposed to AI one perhaps
Intentional typos and deliberately misspelling words.. I have thought... is going to have to become a style. Everyone will have their own way of fudging it up.
Sort of like the painters who came after the "Realism" era.
Interestingly Sam Altman is among those who write without using capital letters -- which LLMs don't do (unless of course specifically prompted to)
Yann LeCun tis saying that some Paul Halpern from X alleges that Albert Einstein (the one from the stories!) had a conversation with some psychologist more than a hundred years ago where those words were uttered.
If true, this is somehow meant to be taken as a rule/law that language and thoughts are fundamentally different things.
No wonder why he got sacked, lol.