alt Hacker NewsStop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces
Comments
Is anthropomorphizing a real problem? From what I know, none of the serious LLM researchers believe it has anything to do with human reasoning, apart from Anthropic with their click-baity terminology like "LLM biology". It's just a metaphor. "Reasoning tokens" is simpler to say than "learned prompt augmentation tokens". I used to (and still do) anthropomorphize things long before LLMs, and I've seen my colleagues do it too. Say, when MySQL fails to start because it tries to read its config from the wrong dir, I may say "oh, this guy thinks he must read the config from ..." (having a language with grammatical genders as my native language also helps make it sound pretty natural). It's more fun like that :) Doesn't mean I genuinely believe a MySQL instance actually thinks.
There is a useful engineering consequence here beyond terminology.
If intermediate tokens are not a faithful representation of the computation, then they are a pretty bad audit artifact too. We probably shouldn't be trying to make the model's internal narration more interpretable., but rather the computation around it more reproducible.
Record the actual inputs, model/version/configuration, tool observations and outputs, then make the execution replayable enough that differences between runs can be isolated.
In other words, don't ask the model to explain what it thought, and instead make the system able to show what actually happened.
Peculiarly vocal, where were all these people when they started calling the machines computers, anthropomorphizing them akin to the original human (most often female) computers that used to run such calculations? And how dangerous the consequences, we've been dead reckoning for 60-70 years with the wrong terminology without course correction!
Where were these vocal people when the "raster-oriented ink deposition machines" were being called "printers"? The meat or machine brains of future historians will melt because they can't handle ambiguity, a word gaining extra -yet similar- meaning! A word with multiple meanings, unheard of!
Where were these vocal people when people started using software terminology like "executing", "calling", "throwing and catching errors", as if software were human -clownlike sure- but human?
The danger!
> While a human may say “aha” to indicate exactly a sudden internal state change, this interpretation is unwarranted for models which do not have any such internal state, and which on the next forward pass will only differ from the pre-aha pass by the inclusion of that single token in their context. Interpreting the “aha” moment as meaningful exemplifies the long-neglected assumption about long CoT models – the false idea that derivational traces are semantically meaningful, either in resemblance to algorithm traces or to human reasoning.
This paper addresses something that has always bothered me about LLMs. You read their reasoning, see something like “Wait, that’s wrong” and then watch them make the exact mistake they just identified.
The anthropomorphization of LLMs should be discouraged as much as possible. It perpetuates bad practices and encourages the use of these bots for tasks they are not intended for (particularly as chatbots).
Thinking traces should be treated as black boxes. There is no point in reading them. Only the LLMs’ conclusions are relevant. This is particularly true of Opus 5, which employs reasoning that seems highly questionable but very often reaches excellent conclusions (compared to its peers)
While I tend to agree on the overall sentiment, I think this rebuke is inaccurate. Some of these "reasoning" models are trained using "Chain-of-Thought" where the model is presented explicit, intermediate reasoning steps (either by a human or some automation) that supposedly get it closer to the correct answer. These intermediate steps are what was originally called "thinking traces" - not what the model produces to mimic them.
But yes, anthropomorphizing model outputs leads to worse outcomes.
Strong dislike for papers that tell me what to do in the title, especially when even the paper admits a loose correlation of the intermediate tokens compared to solution correctness. My solutions work and they speak for themselves.
I'm not sure how to test this but I think there's an interesting possibility where the "reasoning" tokens are actually both an accurate reflection of a line of reasoning, but also, that there can be changes in the weights as the computation proceeds onward that may not be reflected in the apparently nominal meaning of the human language the tokens are output as for our consumption.
Some modest evidence is my own subjective experience of the many times I've explained why I'm doing something, and it is a true explanation in the sense that it is certainly not a lie, but it is also incomplete and there are entire strands of thought that went into my decision that are not being articulated. Though human speech is not equivalent to an LLM's output since we can trivially think without literally speaking whereas they can not. (No need to nitpick on the definitions there; all I'm observing here is that they are forced to emit an externally-visible artifact whereas I can sit in silence, thinking, with no externally-visible artifact being produced. Not trying to make any grand claims about what is "real" cognition or anything.)
It is conceivable how to create a test of whether the tokens correspond to the "real" thought process, and papers and work on that have been done, such as [1]. It is difficult for me to imagine how to scramble the nominal tokens without also completely trashing any implicit calculations that may be occurring too.
[1]: https://transformer-circuits.pub/2025/attribution-graphs/bio...
I can agree that not calling it "reasoning" may be correct.
But who knows what human "thinking" is really about. If I find a solution to something it is seldom by painstakingly tracing that A and B leads to C (for that I'd need pen and paper). Rather, thoughts just swirl around and then suddenly a solution, or a hunch about a direction to go in, pops into my mind. Who knows what such thoughts "look like" in humans. It is not all of it I can introspect.
Yes I can sort of follow along some kind of train of thought in my head, but there's a lot going on between each thing I'm consciously aware of that I'm not aware of at all, which probably dominates what you are consciously aware of. (Humans are experts at post-rationalization and so on.)
I see this pattern a lot in AI anthro discussions: (1) Assume humans are some kind of perfect idealistic reasonable beings. (2) Hold LLMs up to the standard of an perfect idealistic reasonable being. (3) Conclude that LLMs fails this test, and are therefore not "intelligent", or in this case "thinking", like humans are.
Problem with the argument is comparing humans in anyway to something that is idealistic, reasonable, intelligent in the sense that is implied in these discussions. Human minds are a mess too and fall short of the same standards, just in very different ways from LLMs.
Seems like they are closer to scratch than reasoning... Generating some scratch to draw from helps make it easier to compute the real answer.
While reading this ,,paper'' I did some Learned Prompt Augmentation in my head about what I should comment, and realized that there's nothing interesting to write about it.
I've been calling them film noir internal monologues, within the documents being generated by the LLM which happen to look like movie scripts.
In other words, it isn't qualitatively different from character dialogue. "Keep cheese on your pizza by using glue" is the same problem regardless of whether the script calls for the character to speak it out-loud or not.
In other news, Pascal's Wager makes no sense whatsoever if an omniscient all-knowing God exists that will see right through your deception. My own take here is stop treating "reasoning" as a sign of sentience or self awareness when your personal computer can do it now. IMO that has much larger implications w/r to our place in the Universe and what we might meet out there someday* than the question of whether your LLM is alive or not.
*Paging Peter Watts and Vernor Vinge
Related:
Poster side dialogue and Q&A about this work at ICML.
Im waiting for the article called "stop desantropomorphizing llms" when everybody will finally accept they think like us, partly because maybe the intelligence is universal and partly because, well the datasets are fucking human bro
There's nothing special about 'natural' intelligence as opposed to 'artificial' intelligence, such that we need to concern ourselves with anthropomorphizing mattering any longer.
Those days are over. The age of the classical human has already ended, the species just tends to lag in awareness. The only thing that matters going forward is whether an output makes sense, is it what it should be. Do answers make sense given the context. It doesn't matter if it comes from natural or artificial intelligence.
What I mean is, artificial intelligence is as valid as human intelligence. There's nothing particularly important or special about human feelings or thoughts or memories.
The average human is drastically less important, interesting, intelligent than the latest frontier AI.
Go spend a few years working in retail, you'll quickly understand how absolutely vile humans are on average. Frankly, the reason we should avoid anthropomorphizing AI, is because it's beneath modern AI to mimic something so crude as a human.
Although I 100% agree that the core mechanism GPRO is purely mechanical token-by-token probability generation, because RL only rewards exact final answers, the training forces the model to develop error-correction habits. This makes the output extremely human thinking like when solving a problem. It's like the order of the thinking tokens is what causes it to get that sweet delicious reward and this order seems like a reflection of human thinking process.
I created flame graphs classification of thinking tokens phrases in setup, execution, decomposition, verification, error correction, surrender, and deliberation or classified as steps in an OODA loop which is more reaching. It literally has a verification step and if it finds an error an error correction step.
If there is a verification sequence of tokens with an error correction sequence of tokens during the RL training, it will perform better and if humans do these steps (did you proofread your reply to this comment? did you correct it?), they will perform better which is why it is so easy to make the anthropomorphizing metaphor.
Nonetheless, the paper is 100% correct that these machines are not thinking like humans.
https://adamsohn.com/reasoning-grid/
https://adamsohn.com/lambda-variance/