- Author here. Another thing I want to highlight: the language-agnostic "thinking space" finding came from Evan Maunder, who read Part 1 and ran an elegant experiment — same sentence in English, Mandarin, and Base64, cosine similarity at every layer. The representations converge by the early layers, stay nearly identical through the mid-stack, then diverge again at the end as the model commits to an output format.
I extended this to a 2×2 design (two languages × two content types) and the result is even starker: by layer 10, cross-language same-content pairs are more similar than same-language different-content pairs. The model cares about what you're saying, not what language you're saying it in.
This is also what makes layer duplication work — those mid-stack layers operate in a space where input and output distributions match, so you can loop through them without breaking anything. The encoding and decoding boundaries are where the blue walls show up in the heatmaps.
- Author here. The result that surprised me most: after evaluating 3,024 beam search candidates, training a surrogate model on ~4,600 measurements, and scoring 2 million configurations — the Pareto-optimal configs were all simple contiguous blocks. No exotic multi-block compositions, no sparse repeats. Just "repeat layers 31–33" and you're on the efficiency frontier.
I think this says something interesting about how transformers organise computation internally. The mid-stack reasoning circuits are coherent enough that you can loop through them twice without distribution mismatch. The encoding/decoding boundaries are not.
- If you look at convolutional neural nets used in image processing, it's super common for the first layer or so to learn a family of wavelet basis functions. Later layers then do recognition in wavelet space, without that space ever being explained or communicated to the training algorithm.
This work here is obviously more complex than that, but suggests something similar is going on with early layers transforming to some sort of generalized basis functions defining a universal language representation.
- David,
Thanks for this research. I remember being stunned when Goliath showed up and .. worked; this feels like under explored research right now.
I've been thinking about implications of this for local generation -- what's really nice about a repeated layer is it takes up no extra memory -- and therefore works well on the edge.
Can you suggest some exploration angles on the edge side? I've recently started looking at fixing expert layers for an entire generation run as interesting - basically you pay the memory cost once for loading in selected experts - and I think RYS type thinking is a natural extension of this. If you've got some ideas, I'm all ears.
by big_toast
2 subcomments
- This was a little dense for me to grok. Are these well known results or is there an abstract-like summary?
The RYS (repeat yourself) hypothesis that duplicating (the right) layers is enough to improve performance (sorry for not reading closely enough, it's really just stacking the relevant layers?).
The ERD (encoding, reasoning, decoding) layer structure is a relatively robust observation? That the middle layers of the NN will reason in universal space, and this is kinda evidenced by cosine similarities of the hidden states at each layer given similar or dissimilar inputs. And that similar inputs converges by layer 5 and you can kinda watch that happen in the cosine similarities?
This post is incredible and I'm afraid it'll drop off the front page before people engage deeply with it. (The methodology was interesting, maybe there's other big ideas I'm missing.)
by saidnooneever
0 subcomment
- it sometimes makes me think of a video at some point of a guy (Daniel Tammet) who had some brain difference,which caused him to be extremely fast at language learning. He said all language carries the same patterns for him, which he sees through synestesia or whatever.
he learnt icelandic in week and had a fluent conversation on their national TV to prove it. (this is nuts, that language is extremely difficult to pickup with nasal sounds etc.)
ofcourse i guess its not even close to average to have such a abilities as a human, but i wonder if at some point LLMs and AI algorithms and models might shed light on such kind of abstractions (like some mentioned in comments also about image recognition algos) that might help humans actually learn these things themselves, train on them and perhaps even get taught such a thing as a skill.
- Chompsky, in his early days, wrote of deep grammar or deep structure as opposed to surface structure. He theorized that there was a universal set of semantic concepts represented in human brains -- like "verbs" or "noun phrases" and how they interconnect. But Chompsky didn't take this too far.
The work in this paper may be teasing out what this human "deep structure" is, a sort of world view and social view that's at least partly genetically encoded.
- This is orthogonal to quantisation. Could have big impact on smaller models in the 4B-14B range where people often try specific quants and context sizes to fit into the VRAM of a laptop/desktop GPU.
- Relevant:
https://plato.stanford.edu/entries/language-thought/
by notnullorvoid
0 subcomment
- Incredible research. I wonder how close we are to outputting the universal language into it's own reasoning context (which skips encoding layers). Then using the later decoding layers to lazily inspect the reasoning context.
- Fascinating stuff. Any chance of using a sparse autoencoder or some other method to try to grasp what the model is actually doing in those middle layers? It would be quite cool to get a better sense of what type of input it is getting in the first time it goes through the reasoning circuit compared to the second or third time.
by eightysixfour
0 subcomment
- I wonder if there is a path where a model can be trained for variable reasoning layer reuse and determine at a token level how many times to traverse the reasoning blocks. Much like adjustable reasoning levels now, but only repeating thinking circuits instead of running through full output reasoning chains.
While the author mentioned multiple passes through the block didn't help in this instance, I can't help but wonder if it would work if it were built in during training.
- Apologies if I missed this in the article (or in the first article in the series) - what happens if you add two copies of the layer set? Does performance improve over adding one copy of the layer set?
by JPLeRouzic
1 subcomments
- Has anyone started to implement this technique in Llama.cpp or similar inference tool?
- How's the reproducibility of the results? Like avg score of 10 runs vs original.
- [dead]
by OhNoNotAgain_99
0 subcomment
- [dead]
- We've discovered the language. It changes the economics of computing.
As in, this entire cloud buildout is unnecessary because it becomes like using a calculator.
Reach out to chat.
Hacker News