Question is: "Alice has 60 brothers and she also has 212 sisters. How many sisters does Alice’s brother have?" (nb: I have added numbers, it's phrased as X and N in the paper)
I must confess, when I tried to answer the question I got it wrong...! (I feel silly). I only realised I got it wrong when I plugged it into GPT-4o and it came back with the correct answer:
Worth noting that the prompts from the experiment include "To answer the question, DO NOT OUTPUT ANY TEXT EXCEPT following format that contains final answer: ### Answer:" so it appears that they are stopping the models from 'thinking out loud'. If I add that to the prompt, GPT4o gets it consistently wrong...
Also worth noting that there are more complex examples where GPT4o seems to fall down such as:
> Alice has 3 sisters. Her mother has 1 sister who does not have children - she has 7 nephews and nieces and also 2 brothers. Alice's father has a brother who has 5 nephews and nieces in total, and who has also 1 son. How many cousins does Alice's sister have?
However I can't honestly say that this is THAT simple or that most people would get this right...
Note that in the paper, all the numbers used were under 10.
> AIW Variation 1, N=3,M=6,C=7
> AIW Variation 2, N=4,M=2,C=3
> AIW Variation 3, N=1,M=4,C=5
> AIW Variation 4, N=4,M=1,C=2.
Also note that the resricted prompt is only one of the prompt variations tested by the paper. It also explores common techinques to get LLMs to perform better, including "thinking out loud". Even with these methods the models still fail to produce a correct answer.
> Model prompt types. It is well known that so-called prompt engineering can heavily influence the model behavior and model response quality [26, 27, 28]. To account for the response variations due to various prompt forms, we created 3 distinct prompt types asking for the solution to the AIW problem: STANDARD, THINKING, and RESTRICTED. The STANDARD prompt type asks to solve the posed problem and output the final answer in the format as described above. This does not put any specific requirements on model behavior. The THINKING prompt type extends STANDARD with the request to think carefully and double check the solution for any mistakes. This should encourage model to invest more computation into obtaining the solution. In contrast to this, the RESTRICTED prompt urges the model to output only the final answer without any further text. This is supposed to restrict compute invested in producing output. We observe substantially shorter outputs across tested models compared to STANDARD and THINKING for this prompt type (Suppl. Fig. 13).
>Worth noting that the prompts from the experiment include "To answer the question, DO NOT OUTPUT ANY TEXT EXCEPT following format that contains final answer: ### Answer:" so it appears that they are stopping the models from 'thinking out loud'. If I add that to the prompt, GPT4o gets it consistently wrong...
Yes this is a common thing I see people who think LLMs are idiots do.
The more an LLM talks the smarter it gets _because that's the only way it can compute anything_. Imagine saying that Turing machines fail the Church–Turing thesis because they can't solve 3-sat for N variables in N moves or less.
That's what you're doing to an LLM when you ask it to be concise.
As it says in the paper, they already did the research with less limiting prompts. All models were tested with another prompt that gave only the question and no further instruction, and a third prompt that asked the model to consider its answer carefully.
The correct response rate chart doesn't even use the results from the concise prompt.
Even if this were all true, it points to a fundamental risk of using LLM's for important tasks, which is that it is not at all clear to a user that this prompt would cause a problem. The LLM doesn't say "I'm sorry Dave, I just can't do that", it just complies with it and gets the wrong answer.
You can always make excuses for the LLM afterwards, but software with hidden risks like this would not be considered good or reliable in any other context.
While LLMs have incredible potential, and are even downright useful in their current format, they have the rather nasty tendency to confidently present bullshit that passes the smell test for most people. When it's wrong, it's not just wrong, by design, it's wrong but sounds plausible. Considering most people with subject matter expertise aren't going to consult such a bot for their own areas of expertise, that means most people will not be able to detect it intuitively.
Good example: I sunk probably an hour into trying to get Gemini Advanced to help me integrate it with a personal Google Calendar account. I kept asking it things and going crazy because nothing lined up with the way things worked. Finally, it referred to itself as Bard and I realized it was giving me information for a different product. As soon as I asked "are you giving me instructions for Gemini Advanced or Bard?" it was like "OH LOL WOOPS!! YOU GOT ME BRO! XD I CAN'T DO ANY OF THAT! LOL." Which, honestly, is great. Being able to evaluate its answers to realize it's wrong is really neat. Unfortunately, it was neat too late and too manually to stop me from wasting a ton of time.
I have decades of experience working in software-- imagine some rando that didn't know what the hell Bard was or even imagine this thing with "Advanced" in the name couldn't even distinguish between its own and other products' documentation.
> As soon as I asked "are you giving me instructions for Gemini Advanced or Bard?" it was like "OH LOL WOOPS!! YOU GOT ME BRO! XD I CAN'T DO ANY OF THAT! LOL." Which, honestly, is great. Being able to evaluate its answers to realize it's wrong is really neat.
Did it evaluate its answers, or did your expression of doubt cause the eager-to-please language model to switch from "generate (wrong) instructions because that's what the user asked for" to "acknowledge an error because that's what the user asked for"?
How many times have we seen "Oops, you're right! 2 + 2 is actually 5! I apologize for saying it was 4 earlier!"
I suppose whether or not these things evaluate anything at all depends on your perspective, but as soon as I asked if it was giving me the correct information, it did actually give me useful information after that without further prompting.
People really need to stop trying to model an LLM as some kind of magical software component: it all makes a lot more sense if you model it as an under-performing poorly-aligned employee; so like, maybe a distracted kid working for peanuts at your store. You wouldn't trust them to with all of your money and you wouldn't trust them to do a lot of math--if they had to be in charge of checkout, you'd make sure they are only given a point-of-sale terminal and their main job was to, at best, scan the barcodes and compare the total--and yet there are tasks you can imagine handing to them that you'd never give to a robot or computer even though they get it wrong a lot, as not all tasks need to be handled perfectly, they still understand extremely fuzzy tasks, and they are probably cheaper than a qualified adult (certainly cheaper than one who is being paid enough to "give a shit" and pay enough attention to not let you get robbed or even put themselves at some risk for you).
I once gave a 10-dollar bill to a young man serving at the cashier at a store, and he gave me 14 dollars back as a change. I pointed out that this made no sense. He bent down, looked closer at the screen of his machine, and said "Nope, 14 dollars, no mistake". I asked him if he thought I gave him 20. He said no, and even shown me the 10-dollar bill I just gave him. At that point I just gave up and took the money.
Now that I think about it, there was an eerie similarity between this conversation and some of the dialogues I had with LLMs...
Sometimes they get the answer right to something really complex because it fits a pattern, but sometimes they answer with something really really stupid.
People keep saying it because that's literally how LLMs work. They run Montecarlo sampling over a very impressive latent linguistic space. These models are not fundamentally different than the Markov chains of yore except that these latent representations are incredibly powerful.
We haven't even started to approach the largest problem which is moving beyond what is essentially a greedy token level search of this linguistic space. That is, we can't really pick an output that maximized the likelihood of the entire sequence, rather we're simply maximizing the likelihood of each part of the sequence.
LLMs are not reasoning machines. They are basically semantic compression machines with a build in search feature.
> LLMs are not reasoning machines. They are basically semantic compression machines with a build in search feature.
This is just a god of the gaps argument. Understanding is a form of semantic compression. So you're saying we have a system that can learn and construct a database of semantic information, then search it and compose novel, structured and coherent semantic content to respond to an a priori unknown prompt. Sounds like a form of reasoning to me. Maybe it's a limited deeply flawed type of reasoning, not that human reason is perfect, but that doesn't support your contention that it's not reasoning at all.
The best compression relies on understanding. What LLM is is mostly data how humans use words. We understand how to make this data (which is a compression of human text) and use it (generate something). AKA it’s “production rules”, but statistical.
The only issue is ambiguity. What can be generated strongly depends on the order of the tokens. A slight variation can change the meaning and the result is worthless. Understanding is the guardrail against meaningless statement and LLMs lack it.
That's a fascinating insight and it sound so true!
Can you compress for me Van Gogh's Starry Night, please? I'd like to send a copy to my dear old mother who has never seen it. Please make sure when she decompresses the picture she misses none of the exquisite detail in that famous painting.
Okay yes so not really having an artists vocabulary I couldn't compress it as well as someone who has a better understanding of Starry Night. An artist that understands what makes Starry Night great could create a work that evokes similar feelings and emotions. I know this because Van Gogh created many similar works playing with the same techniques, colors, and subjects such as Cypresses in Starry Night and Starry Night over the Rhone. He was clearly working from a concise set of ideas and techniques which I would argue is understanding/compression.
Fine, but we were talking about compression, not about imitation, or inspiration, and not about creating "a work that evokes similar feelings and emotions". If I compress an image, what I get when I decompress it is that image, not "feelings and emotions", yes? In fact, that's kind of the whole point: I can send an image over the web and the receiver can form their own feelings and emotions, without having to rely on mine.
Simple reasoning is a side effect of compression. That is all.
I see from your profile you are focused on your own personal and narrow definition of reasoning. But I’d argue there is a much broader and simpler definition. Can you summarize and apply learnings. This can.
To clarify, what I have in my profile is not my "own personal" definition of reasoning. It's how reasoning is understood in computer science and AI, and I am an expert on the subject through my doctoral studies and my current post-doc research.
That's important to understand. What I have in my profile is not some idiosyncratic idea about reasoning, it's the standard, formal understanding of what reasoning means, as it has developed in practice, in AI research in the last many decades.
I appreciate that there are many people who opine about reasoning who are not aware of that prior work and come up with their own ideas about what "reasoning" means, and some are even AI researches which is very concerning but I can't do anything about that except push back against such uninformed opinions.
Academics have gotten AI wrong since its inception and now are relegated to the trailing edges of the field. Mostly because increasingly insist on theory-as-fact in soft arenas that are clearly still in motion. Reasoning has been one thing, it can continue to grow to be another. But even from your defition, I can provide abductive, inductive, and other examples of it reasoning to this degree just fine. However tour examples are a bit... silly to be honest.
But keep lecturing everyone -- its very common for post-grads to be so up their own behind in their research that they've closed their world off until they are the only ones right in it.
Unfortunately I'm used to people on the internet wearing their ignorance on their sleeve like a badge of honour and so I'm not surprised by the insults in your comment. Just a bit sad to be honest :(
I don't think you can evaluate if an LLM is reasoning by looking purely at the mechanics, because if we looked inside a human brain we wouldn't be able to conclude that it can reason either (our test is 'i think therefore i am', not all these neurons look like they are plugged together in such a way that it enables reason).
It is hard to trust any output from a machine that is confidently wrong so frequently. You need to already be knowledgable in a topic (or at least have a well attuned BS detector) to know if it is giving you correct responses. It can be a time saver and assistant in getting work done where you are already a subject matter expert, but it needs to get better to remove the human from the loop.
No it is because supervised and self supervised learning happen to produce reasoning as a byproduct. For some reason people think that telling a model to recite a trillion tokens somehow will improve it beyond the recitation of those tokens. I mean, in theory you can select the training data so that it will learn what you want, but then again you are limited to what you taught it directly.
The problem is that these models weren't trained to reason. For the task of reasoning, they are overfitting to the dataset. If you want a machine to reason, then build and train it to reason, don't train it to do something else and then expect it to do the thing you didn't train it for.
> The problem is that these models weren't trained to reason.
Except they kind of were. Specifically, they were trained to predict next tokens based on text input, with the optimization function being, does the result make sense to a human?. That's embedded in the training data: it's not random strings, it's output of human reasoning, both basic and sophisticated. That's also what RLHF selects for later on. The models are indeed forced to simulate reasoning.
> don't train it to do something else and then expect it to do the thing you didn't train it for.
That's the difference between AGI and specialized AI - AGI is supposed to do the things you didn't train it to do.
I think people don’t recognize it’s currently doing single turn reasoning and demonstrating the building blocks of real time reasoning with continuous input.
If we tested humans on first thought questions and answers in 5 seconds or less on half the problems we did on LLMs — we might prove humans can’t reason as well
I often want chatgpt to answer concisely and tell it that.
If it really needs to do this 'thinking out loud', could it do that under the hood and not in the final output on my screen? Its first pass could use as many words as it wants to compute the answer, but once the answer is computed please go back and make it short.
Not to take away from your point that maybe the prompt is the problem in these reasoning questions.
I believe the "thinking out loud" is fundamentally part of the process of "text completion" which is what it is doing. Certainly we can (and do) break things apart and add layers that could be used to effectively do this by adding more steps and processing time. But ultimately in a single turn, the entire conversation up to that point (including instructions you may have added telling it to not think out loud) is the input, and the output will reflect that.
>_because that's the only way it can compute anything_
I'm fairly certain we'll soon realize that what's happening here is that the markov chain being run over latent space needs a certain amount of "warmup" before it starts sampling from the optimal region. HMC samplers for Bayesian methods have this same property.
The terms "reasoning", "computing" or "thinking" for this stage should be considered metaphors rather than explanations for what's happening, which is really waiting for a random walk to start sampling from the typical-set.
> Yes this is a common thing I see people who think LLMs are idiots do.
If you're among technologists discussing LLMs academically, as we are, that's a reasonable approach. However, I see a lot of people fail to distinguish that from LLM-powerd products sold to the general public as intelligent bots that can understand your plain english and output answers.
People use their existing mental models when interacting with something. If you have 3 different interfaces with a widget to trigger the same exact function, but one look like a music play button, one looks like a gas pedal, and one looks like mechanical pinball plunger, we interact with those things differently because we know how those things work. In this context, chatbots are designed to engage people's existing mental model for chatting with a person via text. The further you stray from people's expectations of human chat, the further you are from people's expectations, for better, or worse.
If you're selling someone a product claiming it understands plain language questions and gives plain language answers, then not getting the right answer to that question makes it idiotic. The subtleties aren't within most users' grasp, and the "FYI: this thing might be full of shit" disclaimer isn't helpful if you don't know enough about what you're asking to administer a proper smell test.
Your statements are obviously not wrong, but I see people saying these things like its reasonable for non-technical end users to reason about those subtleties. Considering how those things are marketed, I really don't think it is.
I have a blog post in coming on this topic, but yes, this is right.
My method is to first get the LLM to answer the question, and THEN feed the answer back the LLM to extract the answer using constraints + grammar/logit bias/regex to parse the answer. Previously, I constrained to a single true/false token, which worked, but fails on complex queries.
So I split the decision making into a "justification" portion[0], and a "parsing" portion. I found that even crafting the prompt matters here, if you start with or end with, "It's very important to the response includes 'The answer is:'", then the model will lead with that response or only reply with that response. So I put it in the middle of the prompt, and end with with a request to justify the response. As a result, most models will reason their way to the answer, and then end with 'The answer is:'.
There must be a name for the new phenomenon, of which your post is an example, of: 1. Someone expresses that an LLM cannot do some trivial task. 2. Another person declares that they cannot do the task, thereby defending the legitimacy of the LLM.
As a side note, I cannot believe that the average person who can navigate to a chatgpt prompter would fail to correctly answer this question given sufficient motivation to do so.
Many people, especially on this site, really want LLMs to be everything the hype train says and more. Some have literally staked their future on it so they get defensive when people bring up that maybe LLMs aren’t a replacement for human cognition.
The number of times I’ve heard “but did you try model X” or “humans hallucinate too” or “but LLMs don’t get sleep or get sick” is hilarious.
The comment that humans are sometimes wrong too is absolutely ridiculous and I always think about the self-driving car incident in china where the car mistook an advertisement of a car above the road as a real car and hit the breaks causing a chain crash. Even the shittiest driver would never do that.
I like to bring up a certain thought-experiment, where some hypothetical self-driving cars (A) are safe enough to halve the number of the accidents/injuries/deaths compared to human drivers, but (B) almost all of the bad events are from their AI unpredictably deciding to to chase-and-murder pedestrians.
... Would we consider that a good improvement to roll out? Is the only factor short-term utilitarianism?
Sometimes our ability to predict and characterize errors is more important than the total error rate.
Yes. Seems like some users here experience true despair when you suggest that the LLM approach might have a hard limit that means LLMs will be useful but never revolutionary.
Its not just that the LLM fails, its the kind of failures it makes that illustrate how far what the LLM does is from human thinking. Like the problem someone posted where you have a boat and a goat and you want to cross the river, and the LLM's solution is cross the river with the goat, come back with an empty boat, and then cross again with an empty boat.
A human might miss the insight that Alice's brother has one more sister than Alice, namely Alice herself. A human would never think the number of sisters her brother has depends on the number of brothers she has. Not if it has any clue at all what "sister" and "brother" mean.
There's also the phenomenon of papers showing ChatGPT getting it wrong, then people posting anecdotal examples of it getting it right when they try it, but that one already has a couple of names: WORKSFORME or Unable to Reproduce
Hey folks, ever considered the possibility that unreproduceability is not a good thing?
You could call it the "Compose a symphony fallacy", after the scene in I, Robot (2004) where Spooner asks Sonny if he can compose a symphony, to which Sonny replies "Can you?"
Well, why does it need a name? It just seems like a "correct argument". If somebody is claiming that AI is "not all that" because it can make stupid mistakes, surely it's relevant that humans also make stupid mistakes, so AI making them does not necessarily pose a hindrance to it becoming a human-tier reasoner.
It’s worth giving names to interesting phenomena as it helps people to identify and understand them. I find this phenomenon interesting because a person who wants the defend the LLM from apparent criticism is probably more likely to exhibit it.
Even "thinking out loud" we have seen these systems fall flat on their face and give very wrong answers.
Thinking out loud also only gets you so far, if the expectation is a certain type of response it can't always "think out loud". In reality that just proves it isn't really reasoning here and is more likely just self referencing.
That being said, I tried this personally allowing it to think out loud and it told me she has 212 sisters. Using your exact prompt.
> Thinking out loud also only gets you so far, if the expectation is a certain type of response it can't always "think out loud". In reality that just proves it isn't really reasoning here and is more likely just self referencing.
Try to calculate it without writing anything down, or thinking any numbers or words in your head.
You can't draw a 1:1 analogue between an AI and the human experience, but remember that we have an internal stream of consciousness. Maybe the outputs of the LLM are more similar to the stream of consciousness in our heads rather than the words we say? After all, Humans also do lots of self referencing.
> That being said, I tried this personally allowing it to think out loud and it told me she has 212 sisters. Using your exact prompt.
Fair enough, but worst case it can often solve it correctly with the correct reasoning. GPT3.5 can't solve it correctly with correct reasoning, so we are at least appearing to be on a path where AI's can start to solve this question, albeit potentially not fully reliably.
Seems like an odd thing not to state prominently. ChatGPT 4o solved it every time for me.
Seems to work OK even with the "hands tied behind back" prompt. Wonder if this paper is just inaccurate or if OpenAI adjusted the model -- seems unlikely.
Actually, 4o is reported to solve this riddle 60% of the time. I guess I have to test more, but with my about a dozen tests so far, all were correct.
Alice has N Brothers, and she has M sisters.
How many sisters do Alice’s brothers have?
I have not gotten the correct answer to the question as phrased above in one go from Gpt4o yet! (and today was not the first day i tried.)
Phrase it as shown above and you'll likely need 5 or more interactions to get it to generate the correct output. With Gemini i could not get it below 8 without feeling like i was cheating. fwiw.
Chat GPT 4o. I was being a bit generous with background information, but still tests ability to interpret:
------
Me:
Background facts:
Alice is a female human.
All sisters are female, and all brothers are male. No one is their own brother or sister.
Alice has N brothers, and Alice has M sisters.
Now, a few questions based on these facts:
How many sisters do Alice’s brothers have?
Do Alice's brothers have more sisters, fewer sisters, or same number of sisters than Alice has?
Don't forget to consider Alice when counting.
------
ChatGPT:
Let's break down the questions step by step:
How many sisters do Alice’s brothers have?
Alice has M sisters. Since all sisters are female and Alice is included in the total count of sisters for her brothers, Alice's brothers have M+1 sisters. This includes Alice herself and her M sisters.
Do Alice's brothers have more sisters, fewer sisters, or the same number of sisters as Alice has?
Alice has M sisters. Alice's brothers have M+1 sisters (since it includes Alice herself and her M sisters).
Therefore, Alice's brothers have more sisters than Alice has.
In conclusion:
Alice’s brothers have M+1 sisters.
Alice's brothers have more sisters than Alice has.
I don't know the exact encoding and decoding mechanism that ChatGPT 4o has, but I'm pretty sure all the basic facts and rules is already encoded by the models. And you conveniently added the rules that encode the puzzle itself.
The "only output the number" prompt was one of three they used. The others were one with no instructions beyond the question, and one that explicitly told the model to show its reasoning and check for errors.
Of course it's going to give an incorrect answer with that prompt. If the instruction fine tuning is neutered like this prompt, it's going to roll over to the foundation model and offer a completion - probably more influenced by the seed than the prompting text. Bad study.
Edit - I just skimmed the paper - they do use other more appropriate prompt types for reasoning. My initial response was based on the assumption that all prompts used that script prompt quoted in the parent. I retract my "bad paper" comment.
>> I must confess, when I tried to answer the question I got it wrong...! (I feel silly). I only realised I got it wrong when I plugged it into GPT-4o and it came back with the correct answer:
Remember that the authors of the paper did not find that GPT4-o cannot return the right answer. They found that it can't return the right answer more often than ~60% of the time. So you'd have to repeat the experiment many, many times and aggregate the results (the paper uses a binomial Beta this and that etc etc) before you see similar results as the paper.
You won't replicate the results of the paper unless you really put your back into it.
I don’t see such of a distinction between technology and us. We build, drive and continue our overflow this tech. It’s an extension of us inspired by how our own brains work.
One thing that strikes me is that the model first tries using "inclusive language" in one answer - and literally states so, using this specific term - but seems to interpret it in a more mathematical sense (like set inclusion). Then seamlessly switches to the expected DEI spiel in the next paragraph.
For one thing, it makes me suspect that something with the words "inclusive language" was automatically added to the prompt. But more interesting is how it responds to this demand in two different ways, illustrating a "thought process" that is very much unlike that of a human with normal verbal reasoning ability.
I am not a psychologist, but remember reading that schizophrenic people sometimes confuse different meanings of words in a similar way, jumping from one meaning to another without noticing.
> > Question is: "Alice has 60 brothers and she also has 212 sisters. How many sisters does Alice’s brother have?"
> The right answer depends on how Alice identifies I guess? :)
Given that the wording of the question specifically identifies Alice as "she", rather than using a gender-neutral pronoun or no pronoun at all, I think inferring that she identifies as female is reasonable.
Great observation. It would be really interesting to repeat this research with less limiting prompts. I assume they made the restriction to make it easier to parse the answers, perhaps make it two phase, let them respond unlimited, then a follow up prompt ask to compress it to a single answer. I wonder how the results will vary.
> To account for the response variations due to various prompt forms, we created 3 distinct prompt types asking for the solution to the AIW problem: STANDARD, THINKING, and RESTRICTED. The STANDARD prompt type asks to solve the posed problem and output the final answer in the format as described above. This does not put any specific requirements on model behavior. The THINKING prompt type extends STANDARD with the request to think carefully and double check the solution for any mistakes
The thing is that "LLM reasoning breaks down" simply did not surprise me enough that I thought it was worth clicking. Making LLMs fail is not hard. They're interesting for the ways that they work, not the (many, many) ways that they don't.
edit: I've had a look and I don't think any of their prompts are very good. They're certainly not how I'd write them if I wanted a current model to actually solve the problem.
The way to make me take a paper like this seriously would be if you set it up as an adversarial collaboration with a competent prompter, and that person agreed they couldn't make a generic prompt that solved the problem. "We tried three times and none worked" is not news, or at any rate not news about LLMs.
I just tried it with intermediate steps, via the total amount of children Alices mother has, and it worked. So, it seems that an intermediate logical step is too much for these models at the moment, at least in this puzzle where Alice is outside the number given in the description.
I must confess, when I tried to answer the question I got it wrong...! (I feel silly). I only realised I got it wrong when I plugged it into GPT-4o and it came back with the correct answer:
https://chatgpt.com/share/6eb5fa36-e0fd-4417-87d1-64caf06c34...
Worth noting that the prompts from the experiment include "To answer the question, DO NOT OUTPUT ANY TEXT EXCEPT following format that contains final answer: ### Answer:" so it appears that they are stopping the models from 'thinking out loud'. If I add that to the prompt, GPT4o gets it consistently wrong...
https://chatgpt.com/share/7e6a7201-dd2b-43c6-8427-76e5b003ca...
Also worth noting that there are more complex examples where GPT4o seems to fall down such as:
> Alice has 3 sisters. Her mother has 1 sister who does not have children - she has 7 nephews and nieces and also 2 brothers. Alice's father has a brother who has 5 nephews and nieces in total, and who has also 1 son. How many cousins does Alice's sister have?
However I can't honestly say that this is THAT simple or that most people would get this right...