[Gigachad]

The fact that it's looking back and getting confused about what it just wrote is something I've never seen in LLMs before. I tried this on Gemma3 and it didn't get confused like this. It just said yes there is one and then sends a horse emoji.

[Uehreka]

I’ve definitely seen Claude Code go “[wrong fact], which means [some conclusion]. Wait—hold on, wrong fact is wrong.” On the one hand, this is annoying. On the other hand, if the LLM is going to screw up (presumably preventing this is not in the cards) then I’m glad it can catch its own mistakes.

[godshatter]

I wonder what would happen if LMs were built a bit at a time by:

  - add in some smallish portion of the data set
  - have LM trainers (actual humans) interact with it and provide feedback about where the LM is factually incorrect and provide it additional information as to why
  - add those chat logs into the remaining data set
  - rinse and repeat until the LM is an LLM

Would they be any more reliable in terms of hallucinations and factual correctness?

This would replicate to some extent how people learn things. Probably would really slow things down (not scale) and the trainers would need to be subject matter experts and not just random people on the net say whatever they want to say to it as it develops or it will just spiral out of control.

[userbinator]

On the other hand, if the LLM is going to screw up (presumably preventing this is not in the cards) then I’m glad it can catch its own mistakes.

The odd thing is why it would output its own mistakes, instead of internally revising until it's actually satisfied.

[ijk]

So, what I think most people don't realize is that the amount of computation an LLM can do in one pass is strictly bounded. You can see that here with the layers. (This applies to a lot of neural networks [1].)

Remember, they feed in the context on one side of the network, pass it through each layer doing matrix multiplication, and get a value on the other end that we convert back into our representation space. You can view the bit in the middle as doing a kind of really fancy compression, if you like. The important thing is that there are only so many layers, and thus only so many operations.

Therefore, past a certain point they can't revise anything because it runs out of layers. This is one reason why reasoning can help answer more complicated questions. You can train a special token for this purpose [2].

[1]: https://proceedings.neurips.cc/paper_files/paper/2023/file/f...

[2]: https://arxiv.org/abs/2310.02226

[112233]

There is no mechanism in transformer architecture for "internal" thinking ahead, or hierarchical generation. Attention only looks back from current token, ensuring that the model always falls into local maximum, even if it only leads to bad outcomes.

[ijk]

Not strictly true: while this was previously believed to be the case, Anthropic demonstrated that transformers can "think ahead" in some sense, for example when planning rhymes in a poem [1]:

> Instead, we found that Claude plans ahead. Before starting the second line, it began "thinking" of potential on-topic words that would rhyme with "grab it". Then, with these plans in mind, it writes a line to end with the planned word.

They described the mechanism that it uses internally for planning [2]:

> Language models are trained to predict the next word, one word at a time. Given this, one might think the model would rely on pure improvisation. However, we find compelling evidence for a planning mechanism.

> Specifically, the model often activates features corresponding to candidate end-of-next-line words prior to writing the line, and makes use of these features to decide how to compose the line.

[1]: https://www.anthropic.com/research/tracing-thoughts-language...

[2]: https://transformer-circuits.pub/2025/attribution-graphs/bio...

[112233]

Thank you for these links! Their "circuits" research is fascinating. In the example you mention, note how the planned rhyme is piggybacking on the newline token. The internal state that the emergent circuits can use is 1:1 mapped to the tokens. Model cannot trigger an insertion of a "null" token for the purpose of storing this plan-ahead information during inference. Neither there are any sort of "registers" available aside from the tokens. The "thinking" LLMs are not quite that, because the thinking tokens are still forced to become text.

[astrange]

That's what reasoning models are for. You can get most of the benefit by saying an answer once in the reasoning section, because then it can read over it when it outputs it again in the answer section.

It could also have a "delete and revise" token, though you'd have to figure out how to teach it to get used.

[112233]

Given how badly most models degrade once reaching a particular context size (any whitepapers on this welcome), reasoning does seem like quick hack, instead of a thought out architecture.

[captainmuon]

LLMs are just the speech center part of the brain, not a whole brain. It's like when you are speaking on autopilot, or reciting something by heart, it just comes out. There is no reflection or inner thought process. Now thinking models do actually do a bit of inner monologue before showing you the output so they have this problem to a much lesser degree.

[mewpmewp2]

If you did hide its thinking it could do that. But I'm pretty sure what happens here is that it has to go through those tokens for it to be clear that it's doing things wrong.

What I think that happens:

1. There's a question about a somewhat obscure thing.

2. LLM will never know the answer for sure, it has access to this sort of statistical, probability based compressed database on all the facts of the World. Because this allows to store more facts by relating things to each other, but never with 100% certainty.

3. There are particular obscure cases where it hits its initial "statistical intuition" that something is true, so it starts outputting its thoughts as expected for a question where something is likely true. Perhaps you could analyze what it's indicating probabilities on "Yes" vs "No" to estimate its confidence. Perhaps it will show much less likelihood for "Yes", than if the question was for a horse emoji, but in this case "Yes" is still high enough threshold to go through instead of "No".

4. However when it has to explain the exact answer, it's impossible to output an answer because it's false. E.g. seahorse emoji does not exist and it has to output it, previous tokens where "Yes, it exists, it's X", the X will be answers semantically close in meaning.

5. The next token will have context that "Yes, seahorse emoji exists, it is "[HORSE EMOJI]". Now it's clear that there's a conflict here, it's able to see that HORSE emoji is not seahorse emoji, but it had to output it in the line of previous tokens because the previous tokens statistically required an output of something.

[kingstnap]

It can't internally rewise. The last generation produces a distribution and sometimes the wrong answer gets sampled.

There is no "backspace" token, although it would be cool and fancy if we had that.

The more interesting thing is why does it revise its mistakes. The answer to that is having training examples of fixing your own mistakes in the training data plus some RL to bring out that effect more.

[ijk]

There's been a few attempts at training a backspace token, though.

e.g.:

https://arxiv.org/abs/2502.04404

https://arxiv.org/abs/2306.05426

[elliotto]

I do this all the time. I start writing a comment then think about it some more and realize halfway through that I don't know what I'm saying

I have the luxury of a delete button - the LLM doesn't get that privilege.

[VMG]

Isn't that what thinking mode is?

[elliotto]

I tried it with thinking mode and it seems like it spiraled wildly internally, then did a web search and worked it out.

https://chatgpt.com/share/68e3674f-c220-800f-888c-81760e161d...

[drdeca]

AIUI, they generally do all of that at the beginning. Another approach, I suppose, could be to have it generate a second pass? Though that would probably ~double the inference cost.

[godshatter]

If you didn't have the luxury of a delete button, such as when you're just talking directly to someone IRL, you would probably say something like "no, wait, that doesn't make any sense, I think I'm confusing myself" and then either give it another go or just stop there.

I wish LLMs would do this rather than just bluster on ahead.

What I'd like to hear from the AI about seahorse emojis is "my dataset leads me to believe that seahorse emojis exist... but when I go look for one I can't actually find one."

I don't know how to get there, though.

[pixl97]

An LLM is kind of like a human where every thought they had comes out of their mouth.

Most of us humans would sound rather crazy if we did that.

[krackers]

There have been attempts to give LLMs backspace tokens. Since no frontier model uses it I can only guess it doesn't scale as well as just letting it correct itself in COT

https://arxiv.org/abs/2306.05426

[grrowl]

You're describing why reasoning is such a big deal. It can do this freakout in a safe, internal environment, and once it's recent output is confident enough flip into the "actual output" mode.

[Swizec]

> The odd thing is why it would output its own mistakes, instead of internally revising until it's actually satisfied.

Happens to me all the time. Sometimes in a fast-paced conversation you have to keep talking while you’re still figuring out what you’re trying to say. So you say something, realize it’s wrong, and correct yourself. Because if you think silently for too long, you lose your turn.

[catlifeonmars]

That’s probably not the same reason the LLM is doing so though.

[9dev]

Are you sure? Because LLMs definitely have to respond to user queries in time to avoid being perceived as slow. Therefore, thinking internally for too long isn’t an option either.

[rcxdude]

LLMs spend a fixed amount of effort on each token they output, and in a feedforward manner. There's no recursion in the network other than through predicting predicated on the token that it just output. So it's not really time pressure in the same way that you might experience it, but it makes sense that sometimes the available compute is not enough for the next token (and sometimes it's excessive). Thinking modes try to improve this by essentially allowing the LLM to 'talk to itself' before sending anything to the user.

[Sharlin]

There’s no "thinking internally" in LLMs. They literally "think" by outputting tokens. The "thinking modes" supported by online services are just the LLM talking to itself.

[Uehreka]

It’s a lot easier if you (I know I know) stop thinking of them as algorithms and anthropomorphize them more. People frequently say stuff like this, and its pretty clear that our minds process thoughts differently when we directly articulate them than if we act on “latent thoughts” or impulses.

Yell at me all you want about how “LLMs don’t think”, if a mental model is useful, I’m gonna use it.

[catlifeonmars]

It’s only a useful mental model if it leads to useful predictions. Otherwise it’s a “just so” story.

[taylorlunt]

I have a pet theory that LLMs being confused about what they just wrote is why they use so many em dashes. It's a good way to conceptually pivot at any point -- or not.

[Timwi]

I disagree. I don't think LLMs use em-dashes any more than typical professional writing does. It only seems a lot to someone who only reads online chatrooms and social media.

[astrange]

I think the em-dashes are OpenAI's way of watermarking their outputs. Makes it easy to spot spambots on Twitter. Well, that and they all have noticeable favorite words.

[glenstein]

I've seen it several times. He's one where I ask it about baseball teams that don't yet have a hall of fame player:

"Arizona Diamondbacks No inductees wearing their cap.

Randy Johnson is in, but as a Diamondback — so Arizona actually does have one (inducted 2015)."

I asked it to explain what happened and got this, which could be true to how it pulls up info or could be nonsense, not sure:

"Here’s the likely internal sequence:

When you first mentioned the Diamondbacks, my initial retrieval pattern surfaced the “teams without a Hall of Famer” mental list. That list, in my training data, often appears outdated (pre-2015) and still shows Arizona as having none.

A few sentences later, my context pulled up the other knowledge set—Randy Johnson’s induction in 2015 wearing a Diamondbacks cap. That newer, correct fact then overrode the earlier incorrect claim, but because generation is sequential, the earlier error had already been committed to “paper” before the override happened. I corrected mid-stream, but in human terms, it looked like I contradicted myself inside the same paragraph.

This is essentially a knowledge conflict resolution timing issue. I can retrieve and reconcile facts during a single answer, but if the resolution happens after part of the text is already generated, you see the “flip.” It’s not like I “knew it all along but withheld it,” it’s more that multiple memory traces compete, and the most context-relevant one wins—sometimes too late."

[Workaccount2]

Whats fascinating is that these models have excellent knowledge about AI/transformers/LLMs (the labs have clearly been specifically training them in hopes of an automated breakthrough), so they can reason really well about what probably happened.

But it's also just that, what probably happened. They still have no real insight into their own minds, they too are also just victims of whatever it outputs.

[petesergeant]

The inability to do this before was the lack of self-correcting sentences in the training data. Presumably new training corpuses add many more examples of self-correcting sentences / paragraphs?