hckrnws

Author here. I found that duplicating a specific block of 7 middle layers in Qwen2-72B, without modifying any weights, improved performance across all Open LLM Leaderboard benchmarks and took #1. As of 2026, the top 4 models on that leaderboard are still descendants.

The weird finding: single-layer duplication does nothing. Too few layers, nothing. Too many, it gets worse. Only circuit-sized blocks of ~7 layers work. This suggests pretraining carves out discrete functional circuits in the layer stack that only work when preserved whole.

The whole thing was developed on 2x RTX 4090s in my basement. I'm now running current models (GLM-4.7, Qwen3.5, MiniMax M2.5) on a dual GH200 rig (see my other post). Code and new models coming soon.

Happy to answer questions.

The idea that there may be a cognitive lingua franca hiding in the layers is fascinating and gives me hope for a neat idea: pluggable knowledge banks.

MoE notwithstanding, a model trained on the whole Internet and a few hundred thousands stolen books carries way more knowledge than is actually needed for any given workflow. It would be great if we could ship slimmed down models into which we'd plug the knowledge banks useful for today's work, and only those.

It would also mean that you could keep a model's knowledge fresh without retraining the whole of it.

This is interesting. Would this mean less space for hallucination as well (depending on the breadth of knowledge applied to a specific task)?

> pluggable knowledge banks.

plugs in knowledge bank LLM: ... I know kung fu.

I think you may have cracked latent space reasoning. I've had a hunch that something like this would work, but couldn't figure out how the training would back propagate. But you've shown that you just need to duplicate existing layers.

Have you tried a simple inline loop over the duplicated layers? Would be interesting to see performance. Also, would be interesting to compare with a MOE model. See if these layers are acting like different agreeing "experts" or if there is reasoning happening in the latent space.

This is kind of what LoopLM is doing, no? https://arxiv.org/abs/2510.25741

Yes, I've tried duplicating layers, but its not useful.

I think this hasn't been tried before because it's totally unintuitive that feeding the output from later layers into previous ones would actually do anything. And in fact, it usually is detrimental. I guess it takes really bored hobbyists with too much compute to check this stuff.

This works with MoE, and yes, I would be interested in looking into this in more detail. But my wife might disagree with this time sink...

Thanks for the post, really cool stuff you did!

Extra thanks for making it written in a readable and approachable way! I don't have much of a background in this topic, but still managed to understand about 70-80% of it :) You're a good writer

Super cool! Do you do any analysis or have any tools that help you identify these circuits? I came across this [1] recently, and wanted to try to identify specifically strong "circuits" in what seems to be a similar way to what you did.

[1] https://weightwatcher.ai/

I build my own analysis tools. I'm just finishing up running the current generation of LLMs (MiniMax M2.5 and the Qwen3.5 family), and then I will put it all on Github.

It less 'tool', than an assorted set of scripts, tailored to my unusual hardware setup. But it should be easy to extend; I would have released this earlier but I had the (stupid) idea to 'write a paper' on this. Aiming for that delayed this a year. Blogs are the way to go (for me).

Thank you so much for sharing this in a delightful blog post. One of the more enjoyable things I've read in a while. Very motivating!

The dual GH200 build was amazing. Awesome to see someone with such talent & flare in one area also doing great in another area. Thanks for noting that that was you. https://news.ycombinator.com/item?id=46222237

This layer duplication strikes me as a bit of "poor man's" version of looped language models:

https://ouro-llm.github.io/

Pretty cool though. LLM brain surgery.

Agrees, but one thing to note:

I really think from the experiments that 'organs' (not sure what to term this), develop during massive pretraining. This also means maybe looping the entire models is actually not efficient. Maybe a better way is [linear input section -> loop 1 -> linear section -> loop 2 -> linear section -> ... -> loop n -> linear output]?

This would give 'organs' space to develop.

What a great read! You got me at the base64 oddity. I also stumbled over this, while trying to dodge some LLM limitation. (was trying to generate images in a time before multimodal was a thing. it only worked to a degree).

Crazy writeup.

Author is right about the base64 part. Does seem weird that it can decode and understand it at same time. And I guess what makes it weird that we just sorta accept that for say English and German this works ie normal use but when framed as base64 then it suddenly stops feeling intuitive

why tho? it's just an alternate alphabet/set of symbols.

Because its generally expected that models only work 'in distribution', i.e. they work on stuff they have previously seen.

They almost certainly have never seen regular conversations in Base64 in their training set, so its weird that it 'just works'.

Does that make sense?

For all we know, AI tech companies could theoretically have converted all of the "acquired" (ahem!) training set material into base64 and used it for training as well, just like you would encode say japanese romaji or hebrew written in the english alphabet.

Here's an extract, the core TL;DR for a feel of the article.

"And now for the weirdness: There was never the case where any Transformer layer would have seen the output from a future layer!

Layer 10 is trained on layer 9’s output distribution. Layer 60 is trained on layer 59’s. If you rearrange them — feeding layer 60’s output into layer 10 — you’ve created a distribution the model literally never saw during training.

The astounding thing about Goliath wasn’t that is was a huge leap in performance, it was that the damn thing functioned at all. To this day, I still don’t understand why this didn’t raise more eyebrows.

Experimentally, this proved that layers were far more interchangeable than anyone had reason to expect. The internal representations were homogenous enough that the model could digest out-of-order hidden states without collapsing. The architecture was far more flexible than a rigid pipeline.

Between the Base64 observation and Goliath, I had a hypothesis: Transformers have a genuine functional anatomy. Early layers translate input into abstract representations. Late layers translate back out. And the middle layers, the reasoning cortex, operate in a universal internal language that’s robust to architectural rearrangement. The fact that the layer block size for Goliath 120B was 16-layer block made me suspect the input and output ‘processing units’ sized were smaller that 16 layers. I guessed that Alpindale had tried smaller overlaps, and they just didn’t work.

If that was true, maybe I didn’t need to teach a model new facts to make it smarter. I didn’t need fine-tuning. I didn’t need RLHF. I just needed to give it a more layers to think with."

I really enjoyed reading this. I feel like generalists intuitively experience this exact thing so much throughout their lives because they must have this neuroanatomy you describe. There’s a certain geometry to knowledge that makes possible for this orthogonal movement and it is really fascinating to me. Thank you for publishing this, you made my day!

Thanks!

Here is a paper that made a similar observation recently:

https://www.alphaxiv.org/abs/2512.19941

Thanks for the link!

I think that these models have to learn to efficiently use their parameters, and the best way to do that is 'evolve' (yes, a bad word for it), structures over pretraining time. Unfortunately, they don't have a way to access these structures 'from the inside'. I hope this new approach lets up boost performance in s more experimentally rigorous way

I think the recurrence is a consequence of using a residual connection, seems like that makes the representation stay consistent across layers

Very cool, thanks for sharing! Recovering 96% using just two blocks on IMN-1k, wow!

Wild stuff and great read

Do you think karpathy's autoresearch would be useful here?

Based on Karpathy’s writeup the auto research would not have found this. He tells the agent to improve the model and training loop with a five minute time limit, but honestly this “hack” is so far out of distribution that it seems really unlikely an agent would find this.

Is this similar to send 48656c6c6f2c20686f772061726520796f753f in the prompt? As done here: https://youtu.be/GiaNp0u_swU?si=m7-LZ7EYxJCw0k1-

That was a fun read! The base64 decoding and encoding is quite interesting. A parallel: these models are surprisingly robust to heavy word mangling, back in 2023 people used this trick to jailbreak the models very often, but what was more surprising is that they even understand it. I always thought of it this way there must be some circuitry in the model that maps these almost unrecognizable words/sentences into their rectified versions. But what your base64 also shows is the fact thy can also encode them back as well! (However models are known to not be able to produce mangled output that looks convincingly random. I think the base64 transformation is more mechanical in this regard and hence it‘s easier to do the reverse for them.) So your layer circuit hypothesis aligns pretty well with my mental model of how these models work based on the interpretability work I am familiar with! I really also like the way you used the heatmaps as a tool to derive layer insights, very intuitive! But it’s really surprising that you can simply duplicate layers and achieve better results that generalize! This is some research grade effort! I’m confident you could publish this in NeurIPS or ICML if you put it into a paper! I‘m quite impressed! Great work!

Isn't this similar to models that have "double check the answer"?

First pass runs your input through, second pass runs it's output as input?

Just, in double check it presumably runs the entire stack while you're trying to skip the translation steps and only double check the logic?

I don't think its mathematically equivalent or even close because the context/logprobs will be very different, since you only produce 1 token per pass. I'd say the token itself has a lot less information than the signal propagating through the residual stream of transformer blocks.

Maybe, but the interesting thing for me it this only works with specific 'chunks' of the transformer layer stack. More or less that the optimal leads to worse performance.

Really interesting discovery, especially the part about base64. Reminds me of this: Transformer Layers as Painters https://arxiv.org/abs/2407.09298

That's cool. I tried the b64 thing on my local qwen3.5 27b without access to tools and it did it.

Super cool. Love seeing these writeups of hobbyists getting their hands dirty, breaking things, and then coming out on the other side of it with something interesting.

I am not really an ml dev so I don't understand most of it. It does sound ridiculous how it would even work work. Brilliant work and great article I enjoyed reading it

This sounds similar to the Kimi's mixture of experts architecture if I understood it correctly(likely I have not), can you comment on this ?

No worries, happy to discuss anyway :)

MoE (mixture of experts), is an architecture that forces sparsity (not all 'neurons' are active during the forward pass.

This is pretty much orthogonal to that; it works with dense and MoE models, by repeating 'vertical' sections of the transformer stack.

This lines up with what I have seen doing CKA (centered kernel alignment) analysis on transformer internals. The middle layers in most large models have surprisingly similar representations to their neighbors, so duplicating them is basically giving the model extra compute cycles in a region where it is already doing useful refinement without messing up the input/output encoding stages. Curious whether picking layers by representation similarity instead of just a contiguous block would do even better.

Have a look at the boundaries in the heatmaps.

They are of course open to interpretation, but it suggest to me that the models develop 'organs' for processing different types of data, and without duplicating the 'whole organ' you don't get the benefits.

This is quite different to what you usually see, which is via layer ablation experiments. Thoughts?

Maybe you are observing artifacts of Qwen's training procedure. Perhaps they initialized further layers with the weights of previous ones as part of the training curriculum. But it's fun to imagine something more exotic.

This reminds me when people were doing crazy stuff to improve the first Stable Diffusion model by swapping layers, interpolating weights, documenting which layer was most responsible for the quality of the hands etc. At the end the final models had dozens of different ancestors.

Did you ever try multiple copies?

I did, but the combinatorics are mad. I have also tried training a meta-model that predicts the outputs of the combinations.

I will make another post if the topic is popular; its pretty geeky though, even more than my usual blog posts...

very awesome writeup, glad to see someone with access to hw actually playing with this.

Hopefully the cost per GPU will kick-it soon and we'll see people properly play, but frankly the "middle section" layers 2(ish) to (n-1)(ish) of a model can be shuffled up/down and left/right and still perform well.

The fun one will be an LLM router for LLM layers to apply the best reasoning to the best input so far, but frankly that would need the years and years of training that the author hints at.

The one that's still out of grasps is still how to combine/manipulate per-layer k,v caches into a globally coherent state. i.e. if layers can be moved up/down why can't the cached k,v be swapped/combined with different projections? global k,v caches work, but they have to be _huge_ in order to prevent model collapse even on something as simple as owt.

[flagged]

A 5 hour old account with a standard chatgpt reply? Seriously, try harder.