Evaluating chain-of-thought monitorability | Hacker News Reader

Evaluating chain-of-thought monitorability

openai.com

68 pointsmfiguiere21 comments

I might be missing something here as a non-expert, but isn’t chain-of-thought essentially asking the model to narrate what it’s “thinking,” and then monitoring that narration?

That feels closer to injecting a self-report step than observing internal reasoning.

the chain of thought is what it is thinking

Chain-of-thought is a technical term in LLMs — not literally “what it’s thinking.”

As far as I understand it, it’s a generated narration conditioned by the prompt, not direct access to internal reasoning.

It is text that describes a plausible/likely thought process that conditions future generation by it's presence in the context.

Interestingly, it doesn't always condition the final output. When playing with DeepSeek, for example, it's common to see the CoT arrive at a correct answer that the final answer doesn't reflect, and even vice versa, where a chain of faulty reasoning somehow yields the right final answer.

It almost seems that the purpose of the CoT tokens in a transformer network is to act as a computational substrate of sorts. The exact choice of tokens may not be as important as it looks, but it's important that they are present.

IIRC Anthropic has research finding CoT can sometimes be uncorrelated with the final output.

That phenomenon and others is what made it obvious that COT is not its "thinking". I think COT is a process by which the llm expands its processing boundary, in that it allows it to sample over a larger space of possibilities. So its kind of acts like a "trigger" of sorts that allows the model to explore in more ways then without COT. First time I saw this was when I witnessed the "wait" phenomenon. Simply inducing the model to say "wait" in its response improved accuracy of results. as now the model double checked its "work". funny enough it also sometimes lead it to produce a wrong answer where otherwise it should have stuck to its guns. But overall that little wait had a net positive affect. Thats when i knew COT was not same as human thinking as we dont care about trigger words or anything like that, our thinking requires zero language (though it does benefit from language) its a deeper process. Thats why i was interested in latent processing models and foray in that matter.

Wrong to the point of being misleading. This is a goal, not an assumption

Source: all of mechinterp

When we think, our thoughts are composed of both nonverbal cognitive processes (we have access to their outputs, but generally lack introspective awareness of their inner workings), and verbalised thoughts (whether the “voice in your head” or actually spoken as “thinking out loud”).

Of course, there are no doubt significant differences between whatever LLMs are doing and whatever humans are doing when they “think” - but maybe they aren’t quite as dissimilar as many argue? In both cases, there is a mutual/circular relationship between a verbalised process and a nonverbal one (in the LLM case, the inner representations of the model)

The analogy breaks at the learning boundary.

Humans can refine internal models from their own verbalised thoughts; LLMs cannot.

Self-generated text is not an input-strengthening signal for current architectures.

Training on a model’s own outputs produces distributional drift and mode collapse, not refinement.

Equating CoT with “inner speech” implicitly assumes a safe self-training loop that today’s systems simply don’t have.

CoT is a prompted, supervised artifact — not an introspective substrate.

Models have some limited means of refinement available to themselves already: augment a model with any form of external memory, and it can learn by writing to its memory and then reading relevant parts of that accumulated knowledge back in the future. Of course, this is a lot more rigid than what biological brains can do, but it isn’t nothing.

Does “distributional drift and mode collapse” still happen if the outputs are filtered with respect to some external ground truth - e.g. human preferences, or even (in certain restricted domains such as coding) automated evaluations?

I wasn’t talking about human reinforcement.

The discussion has been about CoT in LLMs, so I’ve been referring to the model in isolation from the start.

Here’s how I currently understand the structure of the thread (apologies if I’ve misread anything):

“Is CoT actually thinking?” (my earlier comment)

→ “Yes, it is thinking.”

  → “It might be thinking.”

   → “Under that analogy, self-training on its own CoT should work — but empirically it doesn’t.”

    → “Maybe it would work if you add external memory with human or automated filtering?”

Regarding external memory:

without an external supervisor, whatever gets written into that memory is still the model’s own self-generated output — which brings us back to the original problem.

> Humans can refine internal models from their own verbalised thoughts; LLMs cannot.

can be done without limitations but you won't get the current (and absolutely fucking pointless) kind of speed.

> Self-generated text is not an input-strengthening signal for current architectures.

It can be, the architecture is not the issue. Multi-model generations used for refining answers can also be tweaked for input-strengthening via multi- and cross-stage/link (in the chain) pre-/system-prompts.

> Training on a model’s own outputs produces distributional drift and mode collapse, not refinement

That's an integral part of self-learning. Or in many cases when children raise themselves or each other. Or when hormones are blocked (micro-collapse in sub-systems) or people are drugged (drift). If you didn't have loads of textbooks and online articles, you'd collapse all the time. Some time later: AHA!

It's a "hot reloading" kind of issue but assimilation and adaptation can't/don't happen at the same time. In pure informational contexts it's also just an aggregation while in the real world and in linguistics, things change, in/out of context and based on/grounded in--potentially liminal--(sub-)cultural dogmas, subjectively, collective and objectively phenomenological. Since weighted training data is basically a censored semi-omniscient "pre-computed" botbrain, it's a schizophrenic and dissociating mob of scripted personalities by design, which makes model collapse and drift practically mandatory.

> a safe self-training loop that today’s systems simply don’t have.

Early stages are never safe and you don't get safety otherwise except if you don't have idiots around you, which in money and fame hungry industries and environments is never the case.

> CoT is a prompted, supervised artifact — not an introspective substrate.

Yeah, but their naming schemes are absolute trash in general, anchoring false associations--technically, even deliberately misleading associations or sloppy ignorant ones, desperate to equate their product with human brains--and priming for misappropriation--"it's how humans think".

It is what it is thinking consciously / its internal narrative. For example a supervillain's internal narrative with their plans would go into their COT notepad. If we want to really lean into the analogy between human psychology and LLMs. The "internal reasoning" that people keep referencing in this thread.. referring to the transformer weights and inscrutable inner working of a GPT.. isn't reasoning, but more like instinct, or the subconscious.

It’s more like if the supervillain had to write one word of his chain of thought, then go away and forget what he was thinking, then come back and write one more word based on what he had written so far, repeating the process until the whole chain of thought is written out. Each token is generated conditional only on the previous tokens.

this is not correct

Kind of. The narration is an actual part of the thinking process. Just not the only part.

It can reflect the thinking process fully, or it can be full of post hoc justifications. In practice, it's anything in between.

As task complexity increases and chain-of-thought length grows, it becomes load-bearing by necessity. It still doesn't have to be fully accurate, but it must be doing something right, or the answer wouldn't work.

Related check out chain of draft if you haven't.

Similar performance with 7% of tokens as chain of thought.

https://arxiv.org/abs/2502.18600

That's a comparison to "CoT via prompting of chat models", not "CoT via training reasoning models with RLVR", so it may not apply.

This seems remarkably less safe?

Would would we want to purposely decrease interpretability?

Very strange.

> Our expectation is that combining multiple approaches—a defense-in-depth strategy—can help cover gaps that any single method leaves exposed.

Implement hooks in codex then.

Reply on news.ycombinator.com