May be worth pointing out, that this is not the first residual connection innovation to be in production.
Gemma 3n is also using a low-rank projection of the residual stream called LAuReL. Google did not publicize this too much, I noted it when poking around in the model file.
This is a fantastic catch. I hadn't realized Gemma 3n was already shipping with a variant of this in production.
It feels like we are entering the era of residual stream engineering. For a long time, the standard x + F(x) additive backbone was treated as untouchable. Now, between mHC (weighted scaling) and LAuReL (low-rank projections), labs are finally finding stable ways to make that signal path more dynamic.
I'm curious if the Low-Rank constraint in LAuReL acts as a natural stabilizer against the gradient explosion I saw with unconstrained hyper-connections.
Thanks for the paper link, definitely reading that tonight.
I’m referring specifically to the fundamental residual connection backbone that defines the transformer architecture (x_{l+1} = x_l + F(x_l)).
While the sub-modules differ (MHA vs GQA, SwiGLU vs GeLU, Mixture-of-Depths, etc.), the core signal propagation in Llama, Gemini, and Claude relies on that additive residual stream.
My point here is that DeepSeek's mHC challenges that fundamental additive assumption by introducing learnable weighted scaling factors to the residual path itself.
I guess I am asking how we know Gemini and Claude relies on the additive residual stream. We don't know the architecture details for these closed models?
That's a fair point. We don't have the weights or code for the closed models, so we can't be 100% certain.
However, transformer-based (which their technical reports confirm they are) implies the standard pre-norm/post-nnorm residual block structure. Without those additive residual connections, training networks of that depth (100+ layers) becomes difficult due to the vanishing gradient problem.
If they had solved deep signal propagation without residual streams, that would likely be a bigger architectural breakthrough than the model itself (akin to Mamba/SSMs). It’s a very high-confidence assumption, but you are right that it is still an assumption.
I implemented this for a toy 8M ViT-style model. Got neutral results. This is just an anecdote and is not representative - I think mHC will help with larger parameter sizes and larger token counts.
I suspect your intuition about scale is correct. The theoretical benefit of mHC is that it acts as a sort of relief valve/router for information flow in very deep/wide networks where the standard residual bottleneck becomes an issue. At 8M params, the standard residual stream is likely already perfectly adequate, so mHC might just be adding parameter overhead without solving a real signal propagation problem yet.
Quick question on your run: did you see the signal amplification/instability I saw (values growing during the forward pass)? or was it stable for you, just neutral on loss?
I’ve been wondering for a while: Why isn’t this architecture more common in other LLMs? The context efficiency is amazing, after all - doesn’t that translate to a lot of money at scale?
https://arxiv.org/abs/2512.24880 was published less than two weeks ago, which should explain why it's not more common yet. And it's not that amazing either. It's a slight quality improvement for a slight increase in cost. It's not even clear to me whether it pays for itself.
It's an incremental improvement, not really a revolutionary step.
That being said, I think one could adapt an existing model to add mHC by initializing the routing matrix to the regular residual connection and then post-train the hyper connection matrices. This would let you continue training more efficiently on existing models.
That initialization strategy (effectively starting as identity to match the standard residual stream) is clever. It would let you surgery an existing model like Llama-3 and fine-tune it into an mHC architecture.
The main risk I see is that the 7x signal amplification happens very aggressively. Even with a gentle initialization, you’d likely need very strict gradient clipping or a tiny learning rate on those new routing matrices to prevent them from blowing up the pre-trained features in the first few steps.
Also, I think there's a mix-up here between mHC (this paper, expressivity) and MLA (latent attention, which provides the massive context efficiency). mHC doesn't save memory, but it might make the model 'smarter' per parameter.
Gemma 3n is also using a low-rank projection of the residual stream called LAuReL. Google did not publicize this too much, I noted it when poking around in the model file.
https://arxiv.org/pdf/2411.07501v3
https://old.reddit.com/r/LocalLLaMA/comments/1kuy45r/gemma_3...
Seems to be what they call LAuReL-LR in the paper, with D=2048 and R=64
It feels like we are entering the era of residual stream engineering. For a long time, the standard x + F(x) additive backbone was treated as untouchable. Now, between mHC (weighted scaling) and LAuReL (low-rank projections), labs are finally finding stable ways to make that signal path more dynamic.
I'm curious if the Low-Rank constraint in LAuReL acts as a natural stabilizer against the gradient explosion I saw with unconstrained hyper-connections.
Thanks for the paper link, definitely reading that tonight.
Two key takeaways from the reproduction:
Unconstrained Hyper-Connections really do explode (7x amplification even at 10M scale).
I hit a nasty "stream persistence" bug where my tensors were the right shape, but the architecture was functionally broken.
This is Part 1 (10M scale). Part 2 (scaling to 1B on A100s) is coming later this week. Happy to answer questions about the implementation.
While the sub-modules differ (MHA vs GQA, SwiGLU vs GeLU, Mixture-of-Depths, etc.), the core signal propagation in Llama, Gemini, and Claude relies on that additive residual stream.
My point here is that DeepSeek's mHC challenges that fundamental additive assumption by introducing learnable weighted scaling factors to the residual path itself.
However, transformer-based (which their technical reports confirm they are) implies the standard pre-norm/post-nnorm residual block structure. Without those additive residual connections, training networks of that depth (100+ layers) becomes difficult due to the vanishing gradient problem.
If they had solved deep signal propagation without residual streams, that would likely be a bigger architectural breakthrough than the model itself (akin to Mamba/SSMs). It’s a very high-confidence assumption, but you are right that it is still an assumption.
I suspect your intuition about scale is correct. The theoretical benefit of mHC is that it acts as a sort of relief valve/router for information flow in very deep/wide networks where the standard residual bottleneck becomes an issue. At 8M params, the standard residual stream is likely already perfectly adequate, so mHC might just be adding parameter overhead without solving a real signal propagation problem yet.
Quick question on your run: did you see the signal amplification/instability I saw (values growing during the forward pass)? or was it stable for you, just neutral on loss?
That being said, I think one could adapt an existing model to add mHC by initializing the routing matrix to the regular residual connection and then post-train the hyper connection matrices. This would let you continue training more efficiently on existing models.
The main risk I see is that the 7x signal amplification happens very aggressively. Even with a gentle initialization, you’d likely need very strict gradient clipping or a tiny learning rate on those new routing matrices to prevent them from blowing up the pre-trained features in the first few steps.
Also, I think there's a mix-up here between mHC (this paper, expressivity) and MLA (latent attention, which provides the massive context efficiency). mHC doesn't save memory, but it might make the model 'smarter' per parameter.