But Figure 1 clearly shows that it works, so I don't doubt that it is in fact possible. I'm just struggling to build a picture of how exactly the network accomplishes this.
softmax should be exp()/1+∑exp()
Notice the 1 added to the denominator.
The difference is at the negative limit, softmax can be 0, instead of some epsilon. The same could be done by adding an extra zero value in x.
Downside is, you have to retrain your model from scratch to fix this.
if you think about it, the "escape hatch" is the design of the entire transformer dictionary. if Key/Query attention misaligns with Value's weights, you get a layer head that does not attend to anything...
Still, differential attention is pretty interesting & the benchmarking good, seems worth a try! It's in the same vein as linear or non-softmax attention, which also can work.
Note that there is an error below Eq. 1: W^V should be shape [d_model x d_model] not [d_model, 2*d_model] as in the Q, K matrices.
Idea: why not replace the lambda parameterization between softmax operations with something more general, like a matrix or MLP? E.g: Attention is the affine combination of N softmax attention operations (say, across heads). If the transformer learns an identity matrix here, then you know the original formulation was correct for the data; if it's sparse, these guys were right; if it's something else entirely then who knows...
I didn't follow Miller's proposal quite as he wrote it though and I put the mechanism in all the layers rather than avoiding it at the end.
My test doesn't absolutely rule out usefulness-- there's always different ways of applying something, but I saw no indication of it.
A/B test the two models and compare?
Would be interesting to see if these activations only show up on larger models, or they're some relation to model size.
Hah. Yes. It looks like they only show up in models with 6.7B parameters or more.
The problem can start at 125M. Small enough to test on a whim.
So train a model that exhibits these behaviours, then try it out.
> softmax should be exp()/1+∑exp()
I've also tested it on many networks and different types of attention and I've yet to see a meaningful improvement (or even an improvement), even in generalization.
It really is the training method...
As to the paper, I'm also still at a big lost and honestly, if reviewing could not accept it. The results look good, but I can't tell why and there's some "black magic" going on here.
- Figure 3 has "Transformer" and doesn't specify. Is this StableLM-3B-4E1T?
- What fucking dataset is this on? Stable has a WandB link[2] for that project and I don't see any experiment with similar (presumably entropy?) loss values (come on... this is fucking research... label your fucking graphs...)
- Where the fuck is the ablation? (Yes, I saw Fig 6 and Sec 3.8)
- How do I know that (assuming this is Stable) that the difference isn't just hyperparemeters? Or worse, GPUs! (yes, number of GPUs can change results due to sharding and this changing the statistics)
- How do I know it isn't down to 1k warmup steps instead of 5k?
- What about hidden size, layers, heads, or FFN size? Stable has 32/2560/32/? and this has 28/3072/12/8192 (these all will mess with sharding statistics too). Is the head dimension the same?
- How do I know it isn't down to the tokenizer?
- What is this magic? `0.8 - 0.6 * math.exp(-0.3 * depth)`
- Was this learned? Hand picked? This is a huge factor
- Any information about the learned parameters? Their final values? Trajectories?
- The code does not seem to be the same as whats in the algos...
[0] https://www.evanmiller.org/attention-is-off-by-one.html
[1] This is a bit convoluted but without this condition many "alternative forms" you see would be equivalent to other architectures like linear layers or gated units. Term is not well defined, but this really appears to be the only agreed upon aspect, even if only implicitly stated. This is a much longer conversation though.
[2] https://stability.wandb.io/stability-llm/stable-lm/reports/StableLM-3B-4E1T--VmlldzoyMjU4?accessToken=u3zujipenkx5g7rtcj9qojjgxpconyjktjkli2po09nffrffdhhchq045vp0wyfo
[2.1] The config: https://github.com/Stability-AI/StableLM/blob/main/configs/stablelm-3b-4e1t.ymlI know he's an economist btw. I was also surprised he got a job at anthropic a few months after. I wonder if they're related.
Take a vanilla MHA, tie the V projection between consecutive heads, make the output projection subtract consecutive heads, with some fixed prefactor and voila, you're most if not all of the way there.
In the first diagram with the attention weights, there actually are some negative scores in the noise section. But, the attention to that section is very small anyway. All the second attention map needs to do is predict the noise in the first one -- a task that can be done very accurately, because it has full access to the input of the first.
To refer back to their real-world comparison, noise-canceling headphones have access to what your ear hears through a microphone, so they can output exactly the right cancellation signal. Similarly, the second attention map knows what's being input into the first one, so it can output a corresponding cancellation signal. It's not perfect -- just as noise-canceling headphones aren't perfect -- but it still gets you 99% of the way there, which is enough to boost performance.
I mean, intuitively it would be trivial for the model to just optimise lambda to zero during training. Then you essentially have built a vanilla transformer with an overcomplicated parameter pruning mechanism. Pruning is already pretty well established in the literature as something that works surprisingly good for reducing parameter counts up to (hold on to your papers)... about 40%. In practice the model probably doesn't work exactly like that, but I wouldn't be surprised if it just approximates the normal transformer in the end anyways.
I'm a little concerned about the last sentence of the section introduction of "2 Differential Transformer". It mentions using improvements from previous papers, but in the grammatical context, it's unclear if this improvement is added to both the normal transformer and their diff transformer. This would otherwise sully the comparisons. It's the "main difference" wording in the previous sentence that raised a flag for me.
Of course, a good-faith researcher would know this and may not feel the need to clarify. But you can never be too careful about some published research in this field.
In any event, I'd imagine that this will get widely adopted if the numbers hold up; like I said, this seems to be basically no downside, and should be easy to replicate.
https://github.com/microsoft/unilm/blob/master/Diff-Transfor...
I wonder if the specific setup might be extra effective for coding tuned models as well - you get one coding transformer and one ‘bad habits/chat/other non coding stuff’ negative transformer.
It's also a complex optimization problem, not just about computing. Two times, the parameters take more than two times the time to tune and two times the working memory to train and use. There are also plenty of model training scenarios where data throughput from the dataset into memory and back out is the final bottleneck.
So, though I agree it is indeed a downside, I think it's a worthwhile sacrifice if the results they show are reproducible.
The analogy to noise-cancelling headphones is helpful but in that case we clearly know which is signal and which is noise. Here, if we knew why would we even bother to the noise-cancelling work?
To make things worse, low attention values will have very low gradient, thus needing a lot of weight updates to undo that kind of mistakes. On the other hand, subtracting the output of two softmax allows the model to predict a weight of exactly zero for some of the values, while keeping a reasonable gradient flowing through.
So the model already knows what is noise, but a single softmax makes it harder to exclude it.
Moreover, with a single softmax the output of all heads is forced to stay in the convex hull of the value vectors, whereas with this variant each head can choose its own lambda, thus shifting the "range" of the outputs outside the convex hull pre-determined by the values. This makes the model as a whole more expressive.
Everything I am seeing in this paper is related to reduced size and noise, which implies a reduction in expressiveness.
The improvement in needle and a haystack, benchmarks on multi-hop questions of in corpus data and multishot in-context learning points to this.
This is a wonderful thing if robustness is more important than generality, but it doesn't address trimming away activations that may be spurious in the general use case but may improve an individual domain specificity.
Context would dramatically impact what tradeoffs and more desireble, and noise is probably never desirable. But the ability of this paper to enable bit size for inference points to a reduction in expressiveness.
Perhaps I am too focused on generalization?
Maybe expressiveness is not the right term, or not the main consequence. I could imagine that having different subspaces like that also introduces a degree of robustness to out-of-distribution inputs, as this would make it harder for the outputs of one attention head to shift towards the in-distribution outputs of another head, and thus for the following layer to confuse them.
Also, where is each softmax happening here? For each attention head?
The convex hull (https://en.wikipedia.org/wiki/Convex_hull) of a set is the smallest convex shape that includes that set. Geometrically, it's what you'd get if you "shrink wrapped" the thing you're looking at: edges still protrude, but any indentations get smoothed over.
In this context, the grandparent comment is pointing out that with a traditional transformer block, the resulting computed value for a token can never "stick out" past some weighted average of the values of attended-to tokens, but this differential attention formalism allows that result.
Then y is a convex combination of the v_i, and sits in the convex hull of the v_i.
In the context of standard transformer attention, each output lies in the convex hull ("somewhere between") the input values. With the modification of this paper, the input values can be scaled a little so that the output of different heads can be in different "regions" and thus do not interfere with each other (so yes to your third question, the two softmaxes are performed separately for each head).
https://chatgpt.com/share/67058973-ba94-8008-bed7-c7f9d08dc5...
I just don't see how you could answer these questions without trying it out. And chatgtp DEFINITELY isn't doing that.
Plus the obvious question I'd pose is not in there. What's the difference in performance between this trick and just "softmax() - 0.5 * 2" ? That seems very relevant.
If we’re subtracting one attention matrix from another, we’d end up with attention scores between -1 and 1, with a probability of effectively 0 for any single entry to exactly equal 0.
What’s more, the learnable parameter \lambda allows for negative values. This would allow the model to learn to actually add the attention scores, making a score of exactly 0 impossible.
- Rectify the difference of the softmaxes. (min(0, s(A1) - lambda s(A2)))
- Apply the Heaviside function to the second softmax. (softmax(A1) - lambda H(s(A1) - lambda s(A2))
The second one being a bit more drastic and maybe harder to train.
I wonder if a different approach without that issue exists. For instance, using max(0, exp(x)-1) instead of exp(x) in the softmax attention formula. That way when the query is orthogonal to the key (or worse), it does not contribute.
Won't this cause the gradient to vanish on the left half, causing problems with training?
Wouldn’t this be pretty unlikely, though?
Attention is really good at finding this smattering of words (ie assign most weight there). But it struggles to put exactly 0 on the other words.
SoftmaxA[n] - SoftmaxB[n] is exactly 0?
Even if 2 attention layers learn two different things, I would imagine the corresponding weights in each layer wouldn’t exactly cancel each other out.
The better example is the differential signalling used in professional audio and many digital signaling protocols like Ethernet, HDMI and USB.
Instead of using one wire, referencing to ground, they send the signal as the difference between both wires. Both wires end up carrying the same signal with inverted polarity. Because both wires are running next to each-other any external noice will be applied to both equally.
The voltage will change, but the difference in voltage between both wires is untouched. And when you subtract the two voltages at the receiver end, any noise simply gets subtracted out.
It sort of feels closer to heterodyning and "demodulating" the signal encoded in the softmax. Those tiny little errors we're trying to denoise with this technique are almost closer to carrier waves (when encoded to softmax) than noise imo. This wouldn't get rid of noise in the training data or noise in the dimensionality of the key / value space. It's really only removing noise introduced by the process itself.
The simple way of doing this would be to just remove the softmax or use a sigmoid instead, but in practice a softmax works better it seems.
to eli5:
rope is the modern strategy used to give information to the model about how far a query and a key are apart when doing attention. It's the best strategy we have now, but has a major downside, where it makes some connections between tokens that are far apart much stronger than you would like them to be. Xpos (https://arxiv.org/pdf/2212.10554) is another paper by microsoft tackling issues with rope and you can see figure 1 on page 4 to get a visual interpretation of the sinusoidal attention strength (you would like it to be smooth).
I think a big reason differential transformers is working so well, especially on long sequence stuff, because when both q1 and q2 don't match a token, the rope relative strength will still have the same value and the noise will cancel out. Leaving intended matches, but at the cost of somewhat dampening the original value rope brought.
Just a hypothesis though. It would be easy to test by running this experiment against a baseline where both use alibi attention (https://arxiv.org/pdf/2108.12409) which has a different set of tradeoffs this wouldn't mitigate, but still a really interesting result.
If that is the case, then the "signal" in this case would be the softmax that encodes the dimensions captured by the query / key space. Since the noise ideally is the same in both softmax encodings, subtracting them should "cancel out" the noise.
For example, if you are trying to send a +1V signal on one wire, and a -1V signal on the other and a +0.5V noise exists, one wire will have +1.5V and the other will have -0.5V,
Take the difference and divide by 2:
(+1.5V - -0.5V) / 2 = +1V or, if your setup is different (-0.5V - +1.5V) / 2 = -1V
The analogy I can think of is when you're paying attention to a variety of things and you actively avoid concentrating on something because it will distract you. You don't give it zero attention, you give it negative attention.
If I understand correctly, this architecture trades twice as much attention memory in exchange for either a higher quality model, or less parameters at a similar quality.
> According to the fitted curves, 6.8B-size DIFF Transformer achieves a validation loss comparable to 11B-size Transformer, requiring only 62.2% of parameters
This raises a few questions for me:
- Would having only 60% of the parameters negate the double space for attention, leaving a similar memory profile as a traditional transformer?
- Does that tradeoff change noticeably between training and inference?
Here's the bit from the paper:
> We set the number of heads h = dmodel/2d, where d is equal to the head dimension of Transformer. So we can align the parameter counts and computational complexity.
In other words, they make up for it by having only half as many attention heads per layer.
I wonder about the story behind that formula...
(Although it seems the author do not discuss this choice anywhere in the paper?)
This makes sense, if one considers that the two copies are identical then the softmax outputs would be identical and the difference is zero everywhere. However, by subtracting a scaled copy, the normalization of the difference seems to really boost the signal value(s) over the "noise", making the signal stand out compared to pre-normalization.
softmax(A) - softmax(lambda * A)
I wonder if there's a metaphor here for our own experience and utility in "surprise".
Like if one attention head is surprised by what another learns, up-weight it. But if they both find the same, assume it's not very surprising and down-weight it.
Admittedly, "surprise" is something that has a big section of my knowledgebase[1][2][3] (both as a subjective feeling and adaptive function of our minds, one of the most complex adaptive system we know of)
[1] https://plus.maths.org/content/information-surprise
[2] https://blakeelias.name/papers/Multi-Agent-Cooperation-Intri...
[3] https://complexity.simplecast.com/episodes/81/transcript
I'm wondering if there's any effect of "creativity", or ability to interpolate between concepts. Hallucination and creativity feel very related to me. I understand hallucinating as simply being misaligned with the space humans feel appropriate to interpolate between
Why? I see them as just sampling errors.
Sure a mistake can spark inspiration sometimes, but creativity is much more than mistakes.
> I understand hallucinating as simply being misaligned with the space humans feel appropriate to interpolate between
These language models are next-token predictors. The way the next token is predicted is by sampling a probability space outputted by the model.
That sampling process can be non deterministic.
Hallucinations are when that sampling results in tokens that come together to create a false or otherwise unintended statement.
You can just as well think of everything a model outputs as a hallucination, but we train the model to output a space what we want them to hallucinate is more likely. Otherwise it just outputs meaningless noise.
“Hallucinate” is really an awful word for what it’s trying to describe.
Exactly. Don't forget that an important factor in the success of GPT3 was RLHF, which is essentially training the model to produce "hallucinations" that are more acceptable on average to human trainers.
This is a poor definition that only applies to language models trained to be truthful. If you trained a language model to lie, and it told the truth, that would also be a hallucination.
Or if a model was trained to never sound confident, and it made confident, but correct, claims.
My definition is more accurate.
> Yes the sampling procedure is nondeterministic but this is unrelated to hallucinations.
It’s not the only factor, but it’s absolutely related. It’s also really easy to explain in a comment.
For example, if you always sampled the lowest ranked token, the model would always hallucinate (by output mostly garbage)
Top-k sampling doesn’t eliminate all errors, unless you’re just always picking the most likely token. At that point the sampling process is deterministic, but we’ve seen model output be poor with that setting for reasons I explain next.
> that hallucination is a deeper problem
Of course, it’s because the training process itself is nondeterministic. We can’t make a perfect model, it’s just not how statistical models work.
The model doesn't sample a probability distribution of individual "facts"[1] it samples a probability distribution of tokens which are generally parts of words, bits of punctuation etc. That we get "facts" out of it which may even be wrong in the first place is an emergent behaviour because of the attention mechanism.
Totally agree that it's a deeper problem and may be intrinsic to the design of the models and the fact that they are trained on a next word prediction task. Karpathy talks about the models as "dreaming text". In that sense it's not surprising that some of it is whacky. Our dreams are too.
[1] By which I mean atomic things that can be right or wrong
Hallucination describes the same feature you just called "non deterministic sampling", but exclusively the cases that we don't like. It would be really convenient if we could actually draw that line, but we can't. If non-determinism is a core feature, then that feature will be present in every case; including the ones we find desirable, and the ones we find undesirable.
It looks like creativity has many steps but being able to come with novel, unprompted stuff is important, as long as you are able to discard the bullshit earlier.
"Hallucination" is only a problem if later layers (or additional networks) can't detect and remove it
Yeah I mean sure. Anything is only a problem if it goes undetected. The issue is that if you rely on statistical model, you’ll always have hallucinations, so you can’t filter statistical output with another statistical model if you need real guarantees.
Many products don’t need those guarantees though.
But here’s a case for the other side: sure, most mistakes are just errors, but evolution happens via “mistakes.” Also, LLM’s often deliberately add add randomness at inference time.
That’s a nice slogan, but it’s a gross oversimplification.
In the natural world, you can say that mistakes in DNA replication leads to evolution, but that’s discounting the entire process of natural selection.
Same with creativity. Look at Picasso. His was a technically brilliant realistic painter at 15, but his work later in life evolved to be more abstract and weird. I don’t think that was the result of mistakes, but rather intentionally breaking patterns he learned in his youth.
I don’t know a whole lot about Picasso’s art, but I imagine the way he evaluated his own work played an important role, in being able to see that sometimes creative accidents are interesting.
For one, speed and memory. They have twice as many Q and K weights in the attention blocks, leading to a ~10% reduction in throughput on their H100 (table 7 in appendix A).
Crazy gains though congrats to the researchers
It has to be done in a hierarchical way to know what you attended to + full context.
If the differential vector is being computed with the same input as the attention vector how do you know how to modify the attention vector correctly
They're effectively doing softmax with a fixed temperature, but it's unclear that this work is going to do better than just learning a per-head temperature parameter.
c.f. https://arxiv.org/abs/2010.04245 which shows an improvement by learning per-head temperature.
The other way to think about this is that it looks like a hacked-up kinda-sorta gated attention. If that's the case, then doing softmax(alphaq_1k_1^T - log_sigmoid(betaq_2k_2^T)) might be better? (where alpha,beta are learned temperatures).
Simplified differential T. looks like: (softmax(Q₁K₁) − λ softmax(Q₂K₂)) V
You can factor this into:
x = softmax(Q₁K₁)V
x += -λ softmax(Q₂K₂)V
Then we would know how much this transformer innovation helps by itself.
I'm imagining a smaller model examining the output tokens of a larger model and metaphorically slapping it on the wrist with a ruler if the output tokens start drifting off topic. Not quite the same, but an entertaining thought nonetheless.
But overall... it's mainly a training change, so training is needed to make a difference.
I’m very interested in this claim. I was under the impression that hallucination is unavoidable in these kinds of models. IIRC proof for that was trending on HN a couple weeks ago.
Most of it should be happening when there's no data to draw conclusions from. E.g. STT models make up words in silence, vision models find things in lens cap noise, LLMs make up explanations when they have no data to pull from.
The real solution would be more along the lines of training models to specifically ignore these cases, or in the case of LLMs to just know when to say "I don't know".
edit: not fully but it gives promising results. quiet an improvement actually.
The fundamental issue is that most of the time LLMs are going to be combining statistics derived from many training samples when generating a single continuation, and there is just no guarantee that this will result in a semantically coherent response. Of course the model's depth of parsing and semantic analysis usually means that each generated word is highly plausible, but this isn't the same as being factually correct, especially so in these cases where the model is drawing on multiple sources to create a mashup response, which is the normal mode of operation.
The root problem is simply that the model doesn't capture reality, just an approximation. What we are incorrectly calling "hallucination" is just the best the model has to offer.
during pre-training, there is never an incentive for the model to say "I don't know" because it would be penalized. the model is incentivized to make an educated guess
large transformer models are really good at approximating their dataset. there is no data on the internet about what LLMs know. and even if there were such data, it would probably become obsolete soon
that being said, maybe a big shift in the architecture could solve this. I hope!
Suppose there are many times more posts about something one generation of LLMs can't do (arithmetic, tic-tac-toe, whatever), than posts about how the next generation of models can do that task successfully. I think this is probably the case.
While I doubt it will happen, it would be somewhat funny if training on that text caused a future model to claim it can't do something that it "should" be able to because it internalized that it was an LLM and "LLMs can't do X."
Maybe in the future, those prompts will include motivational phrases, like "You can do it!" or "Believe in yourself, then you can achieve anything."
The guess can be "I don't know". The base LLM would generally only say I don't know if it "knew" that it didn't know, which is not going to be very common. The tuned LLM would be the level responsible for trying to equate a lack of understanding to saying "I don't know"
theres some promising research using this idea, tho i dont have it at hand.
I suppose depending on your point of view, LLMs either can't hallucinate, or that's all they can do.
Empirically, this cannot be true. If it were, it would be statistically shocking how often models coincidentally say true things. The training does not perfectly align the model with truth, but 'orthogonal' is off by a minimum of 45 degrees.
> The training does not perfectly align the model with truth, but 'orthogonal'
Nitpicky, but the more dimensions you have, the easier it is for almost everything to be orthogonal. (https://softwaredoug.com/blog/2022/12/26/surpries-at-hi-dime...) That's why averaging embeddings works.
If you add two vectors that don't have a truth component (ie. are orthogonal to the truth), the resulting vector should be no closer to the truth. If you start with random weights and perform some operation on them such that the new weights have a higher likelihood of producing true statements, the operation must not have been orthogonal to the truth. Am I wrong there?
That's due to the reward function / environment. But even outside extremes like North Korea, lots of education environments value conformity over independent analysis.
Why do you care so much about this particular issue? And why can’t hallucination be something we can aim to improve?
Doesn't an LLM pick the "most probable next symbol" (or, depending on temperature, one of the most probable next symbols)? To do that, doesn't it have to have some idea of what the probability is? Couldn't it then, if the probability falls below some threshold, say "I don't know" instead of giving what it knows is a low-probability answer?
1) The model outputs a ranked list of all tokens; the probability always sums to 1. Sometimes there is a clear "#1 candidate", very often there are a number of plausible candidates. This is just how language works - there are multiple ways to phrase things, and you can't have the model give up every time there is a choice of synonyms.
2) Probability of a token is not the same as probability of a fact. Consider a language model that knows the approximate population of Paris (2 million) but is not confident about the exact figure. Feed such a model the string "The exact population of Paris is" and it will begin with "2" but halfway through the number it will have a more or less arbitrary choice of 10 digits. "2.1I don't know" is neither a desirable answer, nor a plausible one from the model's perspective.
Yes, but that very rarely matters. (Almost never when it's brought up in discussions)
> Couldn't it then, if the probability falls below some threshold, say "I don't know" instead of giving what it knows is a low-probability answer?
A low probability doesn't necessarily mean something's incorrect. Responding to your question in French would also have very low probability, even if it's correct. There's also some nuance around what's classified as a hallucination... Maybe something in the training data did suggest that answer as correct.
There are ideas similar to this one though. It's just a bit more complex than pure probabilities going down. https://arxiv.org/abs/2405.19648
The next bit of confusion is that the 'probability' isn't 'real'. It's not an actual probability but a weight that sums up to one, which is close enough to how probability works that we call it that. However, sometimes there are several good answers and so all the good answers get a lower probability because there are 5 of them. A fixed threshold is not a good idea in this case. Instead, smarter sampling methods are necessary. One possibility is that if we do have seeming confusion, to put a 'confusion marker' into the text and predict the next output and train models to refine the answer as they go along. Not sure if any work has been done here, but this seems to go along with what you're interested in
That's the result after softmax. If you want to act on the raw results, you can still do that.
For one thing the probability of a word occurring is just a probability of the word occurring in a certain sample, it's not an indicator of truth. (e.g. the most problematic concept in philosophy in that just introducing it undermines the truth, see "9/11 truther") It's also not sufficient to pick a "true" word or always pick a "true" word but rather the truthfulness of a statement needs to be evaluated based on the statement as a whole.
A word might have a low probability because it competes with a large number of alternatives that are equally likely which is not a reason to stop generation.
The dot product, which is at the core of attention, is good for similarity not identity. I think this is why models hallucinate - how can they tell the distinction between "I have trained on this fact" and "Looks like something I trained on".
Of course, even if I'm right proper training would account to that by inverting signs where appropriate. Still, it seems weird to present it as the difference, especially seeing as they compare this directly to noise cancelling headphones, where we sum both microphones inputs.
As pointed out by a different comment, it's actually the attention we are interested in that is cancelled out *if they are both equal*. This is what the paper mentions in its abstract;
> promoting the emergence of sparse attention patterns
In theory, it is quite clever, and their results seem to back it up.
But with noise cancelling headphones, we don't sum anything directly---we emit an inverted sound, and to the human ear, this sounds like a subtraction of the two signals. (Audio from the audio source, and noise from the microphone.)
[...] Specifically, we partition the query and key vectors into two groups and compute two separate softmax attention maps. Then the result of subtracting these two maps is regarded as attention scores.
[...] The approach is analogous to noise-canceling headphones and differential amplifiers in electrical engineering, where the difference between two signals cancels out common-mode noise.
Simple change, with seemingly decent improvements across the board.
"The scaling curves indicate that Diff Transformer requires only about 65% of model size or training tokens needed by Transformer to achieve comparable language modeling performance."
"Diff Transformer retains high performance even at reduced bit-widths, ranging from 16 bits to 6 bits. In comparison, Transformer’s accuracy significantly drops with 6-bit quantization. The 4-bit Diff Transformer achieves comparable accuracy as the 6-bit Transformer, and outperforms the 4-bit Transformer by about 25% in accuracy."