News for hackers. Stuff that matters.
Re:Transformers struggle with generalizing tasks beyond pre-... by simbolit on Tuesday November 07, 2023 @12:04PM (#38179441)
Why are transformer models so bad at math? They often fail at simple addition.
Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @12:41PM (#38180045)
Asking a token prediction model to do math is like asking a human to do math without doing the math. What's 9 times 9? I can tell you it's 81 from sheer memorization. I can probably invoke 9 x 9 + 1 = 82 without needing to do any calculation either. But if you ask me 32 * 64, that's very difficult to do without doing calculations. Implicitly doing math is not sensible.
Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @01:02PM (#38180409)
This is wrong. The token predictor will learn algorithms for math.

You can ask GPT-4 arbitrary arithmetic it could never see in it's training set. Even when it's not completely correct, it's extremely close. It is clearly computing algorithms even if those algorithms are not quite right.

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @03:31PM (#38182394)
It is clearly not computing these algorithms. Getting close is more evidence of that than otherwise.

Being able to infer answers to math problems is a thing that humans can do, and that's fine. It's not as good as doing the math though.

Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @03:49PM (#38182629)
It is computing algorithms. There's no debate here.

The question is whether it is computing the right ones.

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @04:06PM (#38182827)
It is not computing the right ones. There is no debate there. These algorithms do not map inputs to outputs reliably.
Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @06:02PM (#38184319)
The point is that this is just blatantly wrong.

>Asking a token prediction model to do math is like asking a human to do math without doing the math. What's 9 times 9? I can tell you it's 81 from sheer memorization. I can probably invoke 9 x 9 + 1 = 82 without needing to do any calculation either. But if you ask me 32 * 64, that's very difficult to do without doing calculations. Implicitly doing math is not sensible.

It's not about memorization. It's trivial to test on instances that would not appear in training and see GPT-4 be better than a human who would attempt the problem without a tool or pad.

The biggest problem with LLM arithmetic is tokenization. https://arxiv.org/abs/2310.02989

Other than that, the algorithms it uses for arithmetic will continue to converge during training until it is correct.

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @07:54PM (#38185359)
I'm not claiming it's about memorization. I have not memorized 9 x 9 + 1 is 82. I can simply do it without invoking any calculation. That is a close equivalent to trying to do with math token prediction.

Humans doing things "without a pad" is not the same thing as doing mental math vs. trying to intuit an answer.

Token predictions will not converge on correct arithmetic algorithms.

Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @09:13PM (#38185921)
>I'm not claiming it's about memorization. I have not memorized 9 x 9 + 1 is 82. I can simply do it without invoking any calculation.

>What's 9 times 9? I can tell you it's 81 from sheer memorization. I can probably invoke 9 x 9 + 1 = 82 without needing to do any calculation

Performing math how you suggest is little more than memorization except you've just memorized a different chain in the process.

>That is a close equivalent to trying to do with math token prediction.

This isn't going to become true no matter how much you repeat it. You've consistently made assertions that are trivially proven false.

>Token predictions will not converge on correct arithmetic algorithms.

Yes it will. It literally will. This isn't some debate. This is something that has been researched.

https://www.alignmentforum.org/posts/N6WM6hs7RQMKDhYjB/a-mec...

You think you have an understanding of Language Models and token prediction. Unfortunately you don't.

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Wednesday November 08, 2023 @06:20AM (#38188921)
> Performing math how you suggest is little more than memorization except you've just memorized a different chain in the process.

It's synthesis of memory perhaps, but it's not memorizing specific answers; which is different yet from executing steps of an algorithm you have memorized.

> This isn't going to become true no matter how much you repeat it. You've consistently made assertions that are trivially proven false.

Then agree to disagree because I think you're full of shit. actually, not even that, I think you just don't even understand what I'm saying. which is agreeing with you 90% of the way. This article is nice. But Memorization -> Generalization -> Cleaning up Noise -> Stability is STILL NOT THE ALGORITHM. It will never be the algorithm. It's always just "generalization-ish". Which is again, a modality that human brains can use to process problems and works pretty well, and can presumably work very well, but is ultimately inferior to established, perfected algorithms.

but nah. you chose petty arrogance so yay, we get get to be dicks. fuck off.

Re:Transformers struggle with generalizing tasks beyond pre-... by criddell on Tuesday November 07, 2023 @01:14PM (#38180569)
If you asked GPT to write a program that multiplies 32 and 64 it could do it.

Since the thing is a computer, why can’t it answer queries by writing and executing a program? Does a transformer-based AI have to be 100% transformer?

Re:Transformers struggle with generalizing tasks beyond pre-... by samus on Tuesday November 07, 2023 @01:25PM (#38180746)
This is simply not how transformer models work. Perhaps overly simplified, it maps an input string to an output string by pushing it through a neural network. But enabling it to access databases, APIs and to write and execute programs on accellerators would indeed significantly increase its capabilities.
Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @03:33PM (#38182410)
Of course that's ok. That's also how brains work, delegating certain tasks to specific specialized parts and processes within the brain.

But it's no longer a token prediction model.

Re:Transformers struggle with generalizing tasks beyond pre-... by criddell on Tuesday November 07, 2023 @04:10PM (#38182883)
If GPT transforms are Turing-complete[1], then why couldn't the ability to write and execute a program be part of a prediction model?

[1]: https://aclanthology.org/2020.conll-1.37.pdf

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @04:41PM (#38183272)
Because the optimization logic of tweaking the gpt transforms via training data and nudging weights and biases to achieve token prediction is not a path to utilizing the full potential of the gpt transforms.
Re:Transformers struggle with generalizing tasks beyond pre-... by criddell on Tuesday November 07, 2023 @05:05PM (#38183620)
I don’t see why not.

I haven’t tried this, but I think I could describe a very basic logical CPU to it and have it execute opcodes and tell me the CPU state at each step (like Knuth’s MIX). Couldn’t I ask it to write programs and execute programs for this computer when doing so would help it give better answers?

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @08:00PM (#38185398)
Highly unlikely to work in the near future. LLMs are bad state machines.

But even if it did, you are basically describing programming in a contrived language, not the llm learning to do math. You can explicitly teach a model to follow an algorithm. But that's not the same as the algorithm being "learned" during training and "invoked" when asked what's 2 + 2?

Re:Transformers struggle with generalizing tasks beyond pre-... by vidarh on Tuesday November 07, 2023 @02:11PM (#38181364)
I halfway agree with this, but not quite.

And GPT4 indeed does just fine with things like 32 * 64 in multiple different ways that humans can also easily memorise the rules for. When I asked it to calculate it step by step, and use easily memorisable shortcuts, it first suggested the "doubling and halving method (though it stupidly started by doubling 32 and halving 64...), and got it right.

I then told it I know the powers of two up to 2^24 by heart, and asked if that changed things.

It then reasonably pointed out this means I know 2^5=32 and 2^6 = 64, and 2^5+2^6=2^(5+6) = 2^11 = 2048 and got the rules right (that was exactly what I intended when I pointed out I remember the powers of two).

So it's not all that awful at these things. It does badly when you effectively try to get it to do maths by blind recall and without nudging it to work step by step, sure.

Where it then falls down tends to be when you ask it to do calculations which involves repetitively applying the same rules many times over, where it will tend to start out well, but occasionally make stupid little mistakes.

If anything the type of mistakes it makes are scarily close to the same kind of lapses in focus humans get when doing the same, where we just get sloppy and fail to add two numbers < 10 correctly for no good reason in the middle of doing it correctly many times, and fail to go back and verify each step.

Where some see LLMs struggling with math, I see LLMs trying to do math in a way that is disturbingly close to how a human school child would, and making the same types of mistakes.

Re:Transformers struggle with generalizing tasks beyond pre-... by spywaregorilla on Tuesday November 07, 2023 @03:38PM (#38182487)
It may or may not be "awful". It's probably much better than humans trying a similar math without math approach.

But it's a clearly suboptimal approach. Humans and AI alike can do well with bad approaches if they must. But we can find alternative ways and we need not shoehorn AI into being LLMs.

Re:Transformers struggle with generalizing tasks beyond pre-... by vidarh on Tuesday November 07, 2023 @03:56PM (#38182704)
We can given them tools, just like we use tools. And I'm sure we can also find ways of "wiring" tools more directly into the models.

I'm not convinced it's better than humans in general at a "math without math approach" yet, but it's certainly better than a lot of people at it. They also do make really trivial mistakes sometimes. But I also don't think there is any indication that any of this is down to things that can't easily be trained out of it. I

Re:Transformers struggle with generalizing tasks beyond pre-... by Verdex on Tuesday November 07, 2023 @12:19PM (#38179698)
I'm definitely not on the LLM bandwagon. While I do think they're interesting, I have significant doubts that they'll be much more than a curiosity to laymen and just one more tool in the box for experts.

However, I'm not sure why being bad at math (if even true based on other comments here), is a legitimate criticism. We've already got a lot of machinery that is very good at math. So, use the tools that make sense for the domain. Or (what I'm really waiting for) better yet find novel ways to stitch the tools that we do have together. Can LLMs turn a problem statement into a matlab program? If so then it doesn't really matter how bad at math they are.

Re:Transformers struggle with generalizing tasks beyond pre-... by vidarh on Tuesday November 07, 2023 @12:17PM (#38179657)
Why are human children so bad at math until they've repetitively trained on how to apply each basic operation over and over and over even after having learned rules that should be simple to apply?

As it happens, conditioning even people to actually apply rules properly tends to take a lot of repetition. How many individual examples of step-by-step working out basic math problems do you think they have been in their training data?

Prompt them the way you'd prompt a child who is sloppy into working step by step and explaining how it applies the rules it has learned, and it will tend to do better. While tokenization might not help, I don't think there's an inherent problem there beyond feeding them enough training data. Whether that's worthwhile vs. having it resort to tools, is another matter.

Re:Transformers struggle with generalizing tasks beyond pre-... by haltist on Tuesday November 07, 2023 @12:09PM (#38179536)
Neural networks approximate the statistical distribution of whatever they are trained on. There is no reason to expect that any neural network will perform well on out of distribution tasks, e.g. multi-digit addition and multiplication.
Re:Transformers struggle with generalizing tasks beyond pre-... by ctoth on Tuesday November 07, 2023 @12:12PM (#38179577)
This is (probably) wrong, see:

https://gwern.net/scaling-hypothesis And specifically the part where it discusses addition under the heading Blessings of Scale.

Re:Transformers struggle with generalizing tasks beyond pre-... by haltist on Tuesday November 07, 2023 @12:16PM (#38179636)
So why can't GPT do arithmetic? At what scale do you suppose it will manage to add numbers?
Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @12:34PM (#38179943)
GPT-4 can do arithmetic just fine. In fact, despite being severely hobbled by Tokenization (https://arxiv.org/abs/2110.09485), it can perform arithmetic better than you could without a dedicated tool or pad.
Re:Transformers struggle with generalizing tasks beyond pre-... by Der_Einzige on Tuesday November 07, 2023 @12:42PM (#38180060)
Besides your answer on tokenization being right on the mark (gwern even has articles discussing this exact issue - https://gwern.net/gpt-3#bpes), that linked paper is extremely powerful for theoretical thinking about this problem. I literally just left a comment claiming that almost all of human thought is encoded in LLMs, leading almost all tasks to be considered interpolation.

If this is true, I suppose based on what we know of the curse of dimensionality it makes sense. Very thought-provoking work.

Re:Transformers struggle with generalizing tasks beyond pre-... by og_kalu on Tuesday November 07, 2023 @12:49PM (#38180193)
Damn I linked the wrong paper lol.

Meant to link this. More pertinent to my point on Tokenization https://arxiv.org/abs/2310.02989

Re:Transformers struggle with generalizing tasks beyond pre-... by craftit on Tuesday November 07, 2023 @12:14PM (#38179615)
Because often, the tokens are broken up as random groups of numbers. For example, let's say 1984 appears quite a few times in the source text, this will become a single token. Given that these many different, semi-random groups of digits it is hard for the LLM to learn any consistent rules. I believe there are papers showing that if you structure numbers more consistently LLMs have no problem with this kind of arithmetic.
Re:Transformers struggle with generalizing tasks beyond pre-... by ffwd on Tuesday November 07, 2023 @04:15PM (#38182946)
I have a question which I don't know the answer to:

With those structured numbers will the LLMs be 100% accurate on new prompts or will they just be better than chance (even significantly better than chance)?

Because this is one thing, it has to learn the structure and then create probabilities based on the data, but does that mean it's actually learning the underlying algorithm for addition for example or is it just getting better probabilities because of a narrowing of them? If it can indeed learn underlying algorithms like this that's super interesting. The reason also this is in an issue if it _can't_ learn those, you can never trust the answer unless you check it, but that's sort of a sidepoint.

Re:Transformers struggle with generalizing tasks beyond pre-... by craftit on Thursday November 09, 2023 @03:58AM (#38202632)
From what I understand, it can learn and execute the algorithm fairly reliably, though it won't be 100%. When the LLM generates text, it is randomised a little, as well as some tricks that prevent repetition, which would likely cause problems with numbers containing all the same digit.
Re:Transformers struggle with generalizing tasks beyond pre-... by tsunamifury on Tuesday November 07, 2023 @12:10PM (#38179537)
They aren't. This is overblown if you ask them to do simple math alone they are fine. And what they can't do they can do better with reflexion. It's not perfect, but they aren't terrible.
Re:Transformers struggle with generalizing tasks beyond pre-... by simbolit on Tuesday November 07, 2023 @12:12PM (#38179591)
Just one (of many) examples: I asked for scores in several different categories, and the total score was not what the individiual scores added up to. I mentioned this, the model apologized, and gave another total score (also wrong).
Re:Transformers struggle with generalizing tasks beyond pre-... by ctoth on Tuesday November 07, 2023 @12:10PM (#38179545)
Tokenization, same as a year ago.
Re:Transformers struggle with generalizing tasks beyond pre-... by GaggiX on Tuesday November 07, 2023 @12:10PM (#38179538)
If you give them a calculator they will be much better.

A better answer: every time the model needs to predict the digit of a sum the model needs to solve the entire addition to know the carries, a bigger sum requires a bigger model to solve them.

Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Tuesday November 07, 2023 @12:07PM (#38179490)
Don't they essentially memorize all examples they've seen, which contrasts with understanding how to solve problems as in addition with carries on digits etc?
Re:Transformers struggle with generalizing tasks beyond pre-... by zamadatix on Tuesday November 07, 2023 @12:20PM (#38179716)
This oversimplifies too much. GPT4 is able to add random large numbers too large to have been in the data set. When it's wrong it tends to be closely wrong e.g. maybe the first 4 digits of division on a large number by a decimal number with many fractional digits are correct. Neither of these examples can be explained by memorization alone, other operational mechanics are being applied to get these outcomes.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Wednesday November 08, 2023 @11:57AM (#38193237)
To your point, it would be really neat if we could trace what it's doing to get answers, see why it's sometimes right, sometimes close but wrong!
Re:Transformers struggle with generalizing tasks beyond pre-... by zamadatix on Wednesday November 08, 2023 @02:29PM (#38195656)
You can trace it pretty easily in a local neural net which takes a couple of minutes to train. The only reason we can't trace it in GPT is it's a closed model so we don't have access to do so.

Largely two things come into play: 1) Some part of the neural net is emulating more traditional logic, but it may not always be the most activated part or tuned to be perfect in the answers 2) There isn't really a "jmp" equivalent in a single iteration, so the neural net has to learn to not only do decimal division but do it based on iterating output tokens, continuing perfectly each token output, and choosing to put the right stuff into context and keep that context activated at the right time.

"Activated" in this case means, more or less, the group of neurons specializing on this task are being both fed and listened to.

You can even train a neural net to emulate a traditional addition circuit directly, it's just less efficient than one would think if you're trying to build a general purpose model instead of a specialized one.

Re:Transformers struggle with generalizing tasks beyond pre-... by tsunamifury on Tuesday November 07, 2023 @12:10PM (#38179549)
No, stop repeating this. They use examples to learn and attempt to generalize with on-the-fly modeling.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Tuesday November 07, 2023 @12:37PM (#38179988)
Citation instead of handwaving that they can learn algorithms? Also, I’ve not asked that before so if it’s being asked frequently, I would imagine it might be because it’s a good question.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Tuesday November 07, 2023 @12:14PM (#38179617)
Are you claiming the on the fly modeling can learn or reproduce multistep algorithms, like simple addition that the parent comment asked about, and which i asked about above?
Re:Transformers struggle with generalizing tasks beyond pre-... by vidarh on Tuesday November 07, 2023 @01:28PM (#38180779)
I just fed it this:

---

Here is a grammar for a made up language:

    sentence = "sob" subject (verb "ob" object | verb "ad" sentence | "ploj").
    verb = "ab" | "pom"
    object = "foo" | "bar"
Construct a few sentences, and show the sentence so far, and which rule and step of the rule you are currently applying. E.g. sentence= ""; rule=sentence; step="sob".

---

Notice I did not explain the BNF syntax, which is reasonable as it'll have plenty of that in its training data. Notice that I also "forgot" to give it a production for the "subject" rule - leaving it with two choices: It can give up and tell me off, or it can do as best it can and let me know it's had to improvise.

Here's the first sentence construction:

---

First Sentence Construction:

sentence = "";

Rule: sentence; Step: "sob". sentence = "sob";

Rule: subject; Step: (choosing object as a subject for simplicity) "foo". sentence = "sob foo";

Rule: verb; Step: "ab". sentence = "sob foo ab";

Rule: (verb "ob" object | verb "ad" sentence | "ploj"); Step: "ob". sentence = "sob foo ab ob";

Rule: object; Step: "bar". sentence = "sob foo ab ob bar"; (Complete Sentence)

---

It has correctly applied the rules I did give, and opted to "fill in" "object" as a replacement for my missing "subject" rule. There are few enough rules this could be a fluke, but it's a better choice to guess at that substitution than verb or sentence, though hardly takes a genius.

Second:

---

Second Sentence Construction:

sentence = "";

Rule: sentence; Step: "sob". sentence = "sob";

Rule: subject; Step: "bar". sentence = "sob bar";

Rule: verb; Step: "pom". sentence = "sob bar pom";

Rule: (verb "ob" object | verb "ad" sentence | "ploj"); Step: "ad". sentence = "sob bar pom ad";

Rule: sentence; Step: "sob" (this starts a nested sentence). sentence = "sob bar pom ad sob";

Rule: subject; Step: "foo". sentence = "sob bar pom ad sob foo";

Rule: verb; Step: "ab". sentence = "sob bar pom ad sob foo ab";

Rule: (verb "ob" object | verb "ad" sentence | "ploj"); Step: "ob". sentence = "sob bar pom ad sob foo ab ob";

Rule: object; Step: "foo". sentence = "sob bar pom ad sob foo ab ob foo"; (Complete Sentence)

---

This is also correct, and it correctly points out that the arm it picked triggered a nested application of the `sentence` rule.

It stumbled on the third attempt:

---

Third Sentence Construction:

sentence = "";

Rule: sentence; Step: "sob". sentence = "sob";

Rule: subject; Step: "foo". sentence = "sob foo";

Rule: verb; Step: "pom". sentence = "sob foo pom";

Rule: (verb "ob" object | verb "ad" sentence | "ploj"); Step: "ploj". sentence = "sob foo pom ploj"; (Complete Sentence)

---

Note that there is a possible interpretation of the "|" operator which would make its sentence valid, so "technically" it could be, but it would involve (ver "ob" object | verb "ad" sentence | "ploj") to be interpreted as (verb (("ob object)|("ad" sentence)|"ploj")) and that's contrived.

So to answer you: It can, and it does, but not always consistently. One can choose to be disappointed it's not perfect, but for just handing it a grammar and asking it to reason about it with no example text or chances to "practice" whatsoever, I think it's doing quite well.

I've interviewed many software developers who'd struggle with a task like this.

When prodded about the third example, it picks a more reasonable interpretation of the "|", though it's explanation starts out with an error, but then given a correct explanation:

---

Q: How do you figure this? The full rule is "sob" subject (verb "ob" object | verb "ad" sentence | "ploj"). How do you figure the verb is allowed without being followed by either "ad" or "ob"?

A: Upon closer examination, you are correct to point out that the structure following "sob" subject must indeed involve a verb according to the rule provided. The full rule:

"sob" subject (verb "ob" object | verb "ad" sentence | "ploj"). indicates that the sentence must start with "sob", followed by a subject, and then either:

a verb followed by "ob" and an object,

a verb followed by "ad" and another sentence,

or the standalone term "ploj".

---

It goes on to give two correct alternatives to the erroneous one:

---

* "sob foo ploj" — where "ploj" follows directly after the subject without a verb, or * "sob foo ab ob bar" — where "ab" is the verb followed by "ob" and "bar" as the object.

---

While it's not perfect, it can very clearly "learn and reproduce multistep algorithms" to some extent. It also clearly needs to get better at it.

Re:Transformers struggle with generalizing tasks beyond pre-... by tsunamifury on Tuesday November 07, 2023 @09:42PM (#38186115)
Yea I mean you don’t even have to go this far but it’s obvious it can do logical, mathematical and systematic reasoning. I have no idea why people keep insisting it can’t.
Re:Transformers struggle with generalizing tasks beyond pre-... by vidarh on Wednesday November 08, 2023 @05:02AM (#38188470)
Yeah, I asked it that partly because I got curious if it could handle a task that well (because I know plenty of people who can't...), and partly because so many people keep rejecting other reasoning tasks.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Wednesday November 08, 2023 @12:06PM (#38193417)
People keep asking because they are looking for a rigorous example in a citation that's unambiguous. If it's not obvious, that is even more compelling.

I personally think the more theory trained folks wonder if there will turn out to be some analogy with the corresponding towers of languages and machines recognizing them (regular languages, finite automata, push down automata, turing machines, for examples).

https://a.co/d/8ix6V6m

is an example covering the historical language classes and recognizers for them.

Perhaps LLMs and various architectures will turn out fitting what you describe.

Re:Transformers struggle with generalizing tasks beyond pre-... by GaggiX on Tuesday November 07, 2023 @12:10PM (#38179552)
No they don't.
Re:Transformers struggle with generalizing tasks beyond pre-... by ctoth on Tuesday November 07, 2023 @12:13PM (#38179608)
Yes, actually, they do. See the heading Blessings of scale in [0]

[0]: https://gwern.net/scaling-hypothesis:

Re:Transformers struggle with generalizing tasks beyond pre-... by GaggiX on Tuesday November 07, 2023 @12:23PM (#38179761)
I have trained models on performing additions on numbers with fixed number of digits, just to prove to a friend that a neural network can learn to do addition, I hide in the dataset many combinations of numbers but the model was still able to sum them correctly, therefore it learned to perform addition on numbers with fixed number of digits. So no it's not memorizing, it's something that I have tested myself.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Tuesday November 07, 2023 @12:27PM (#38179824)
Interesting. Thanks for the comment. What if you ask it to add numbers with more digits?
Re:Transformers struggle with generalizing tasks beyond pre-... by GaggiX on Tuesday November 07, 2023 @12:38PM (#38180005)
The model is trained with a fixed number of tokens, I think it would cause problems with the sinusoidal embeddings layer.
Re:Transformers struggle with generalizing tasks beyond pre-... by Der_Einzige on Tuesday November 07, 2023 @12:46PM (#38180132)
Fine-tuning the model to force it to understand the insane tokenization its forced to use for its vocabulary and perform accurate addition sounds great until you start trying very large numbers or messing with whatever decoding settings you're using (and let me guess, it was beam search).

Or, you can simply ask it to use a tool, like a calculator for you. This is more reliable than fine-tuning a current technique BPE tokenized model ever will be.

Re:Transformers struggle with generalizing tasks beyond pre-... by GaggiX on Tuesday November 07, 2023 @05:56PM (#38184242)
I trained my models from scratch, the tokenization is extremely simple and designed just for this task, decoder-only models can learn to do sums of arbitrarily large numbers if the model is large enough (they are computationally bounded). The decoding method used was greedy or simple sampling.
Re:Transformers struggle with generalizing tasks beyond pre-... by jjtheblunt on Wednesday November 08, 2023 @01:27PM (#38194692)
super cool and interesting link. thanks.