Would be interesting to combine it with Reasoning In the Latent Space: feed the vector from the output layer of transformer back to input.

Obviously, you can't do it in pre-training. But you can add it later as an optional 'extra' vector, I think. E.g. `input_embedding + MLP(prev_output) * alpha`. Alpha is zero during pre-training.

I like this plan, but don't you already have this from the input vector in the prompt, at least if the inference is 'chunk wise' - generating a latent space vector, decoding it, outputting it, doing the next one.

What if you trained a separate thinking phase using the auto encoder, though? Might be more efficient, and then you've got it using neuralese internally.

Actually, reading the (summary) paper - they tried your idea and had trouble with it for a different reason:

   > Once the generative head predicts the next vector , a natural next step would be to feed it directly as input to the Transformer for predicting . However, we found that the model struggles to unpack the semantic information from such a compact representation. Instead, we ground the autoregressive process back in the more structured discrete space, where the predicted  is passed through the autoencoder to reconstruct the K tokens.

Very interesting. Also I find these training parameters quite elegant:

- Diversity: This term encourages the model to generate a diverse set of samples, preventing mode collapse. - Fidelity: This term rewards the model for making predictions that are close to the ground-truth

I'm wondering if a continuos next-vector generative approach also increase innate "reasoning" capabilities of the model, since it could potentially capture more of the semantics of the data vs just tokens.

And may be even more adapted to sorts of RL finetuning?

They say this technique isn't compatible yet with RL because you can't adjust the logits. So no GRPO I guess, which is going to be the biggest issue. An LLM with no RL applied isn't going to be that useful.

If they can reinvent RL so it works with this then I guess the big labs will be all over it, as ~halving inference costs would be huge (especially if Ed Zitron's leaked OpenAI inf costs are accurate). Potentially the difference between inferencing being profitable and loss making. It's an elegant approach.

I also wonder how far they can push K if other aspects are tweaked. The approach of just doubling each parameter each time leaves a lot of space between the chosen value and the next value known to not work.

If this works, we’re looking at the next structural shift in LLMs — and all the “bigger model = better” business might finally face a serious challenger. But — and you knew there’d be a “but” — if the reconstruction fails in edge-cases, or the continuous space hides weird failure modes, then this could backfire and produce models that look efficient but feel brittle.

Still — props to the team for going after the real root of inefficiency, not just piling on more layers. If nothing else, this is one to watch if you care about scaling models smarter.

The technique of compressing tokens down reminds me a bit of byte latent transformers

K being fixed here seems like it will eventually be done away with

When I'm thinking about math proofs, sometimes I can have a single idea which can be unfolded into a hundred lines of proof

Maybe I'm getting the wrong analogy here, but if vectors = ideas then K should depend on the vector

Congratulations for the authors, but damit, there goes a good idea ^^