Hi everyone,
I’ve just published a new preprint that brings together several months of experiments on hidden-state dynamics in small open Transformer models.
The question is fairly simple:
During inference, do internal representations simply change from layer to layer, or is there evidence of a more structured progression across depth and generation time?
I tried to study this without assuming that hidden-state dynamics are equivalent to “reasoning”.
The working framework is:
tokens → embeddings → contextualisation → relational structuring → functional structuring → decision formation → projection
This is a descriptive hypothesis about representation dynamics, not a claim that these stages correspond to a universal reasoning mechanism.
The expanded study uses 8 locally instrumented open models, with synchronized hidden-state and output observations and explicit separation between:
depth — what changes as information passes through Transformer layers
time — what changes as autoregressive generation progresses
A few results were particularly interesting.
First, local ordering across model depth survived expansion.
The observed ordering was significantly more structured than random layer permutations (p = 0.00019996
) and remained supported when each model was removed from the panel one at a time (8/8 leave-one-model-out checks). Second, cross-model depth profiles remained surprisingly coherent.
The mean correlation across normalized depth profiles was approximately r = 0.789.
This does not mean that all models follow the same trajectory. Rather, it suggests that some aspects of where changes occur along depth may be more shared than I initially expected.
Third, functionally labelled events were not uniformly distributed across depth.
Event type showed a statistically supported association with normalized layer depth (p = 0.0024
).
I’m deliberately calling this an association, not evidence of a causal mechanism.
But one of the most useful results was actually a failure to replicate.
In an earlier smaller panel, a common temporal pattern in local trajectory instability looked promising. After expanding the panel, that common temporal mode disappeared — it survived 0/8 leave-one-model-out checks.
Two other intuitive hypotheses also failed:
models with similar observed functional outcomes were not significantly more structurally similar (p = 0.408
), and models from the same architecture family were not significantly more similar either (p = 0.771
).
To me, this is probably the most important part of the result.
The data do not support a simple story where architecture determines one characteristic trajectory or where one universal temporal dynamic explains inference.
What remains is a narrower hypothesis:
Transformer inference may contain reproducible structure along depth while remaining highly conditional in time and behavior.
I refer to this as Progressive Representational Structuring.
The framework is summarized by:
Representation ≠ Function ≠ Behavior
A representation can contain information without that information yet serving the same function, and a functional transition does not guarantee a particular final behavior.
I would be especially interested in feedback from people working on:
mechanistic interpretability, activation patching, probing, hidden-state geometry, steering, representation engineering, or larger open models.
In particular, I’m curious whether others observe similar **ordered depth structure without a universal temporal trajectory.
Preprint:
Progressive Representational Structuring in Small Language Models: Functionally Labelled Trajectories Across Depth and Time
DOI: 10.5281/zenodo.22116637 This is still descriptive work. Causal intervention and structural-transfer experiments are separate next steps rather than claims of this paper. Progressive Representational Structuring in Small Language Models: Functionally Labelled Trajectories Across Depth and Time | Zenodo