THEORY-067: A transformer can run a learning algorithm on a model held in its activations, and the architecture admits several, so a construction identifies none

The account

Write the in-context examples into the token stream as (x_j, y_j) pairs. Then some of the network’s activations can be read as the parameters of a small model, and a layer can be read as an update to those parameters. On this view a forward pass over a prompt is a training run over a dataset, and the prediction for the query token is the trained model’s output.

The frame is not an analogy. It has been made precise three times, for three different algorithms, by three groups working independently.

What is established

• One gradient step per layer. LIT-533, Proposition 1: with W_K = W_Q = [[I_x, 0], [0, 0]], W_V = [[0, 0], [W_0, −I_y]] and P = (η/N)I, a linear self-attention step on every token is exactly the token change (0, −ΔW x_j) that one gradient-descent step on the regression loss induces.
• A closed-form ridge update. LIT-532, Theorem 2: a transformer predicts according to a single Sherman–Morrison update with a constant number of layers and O(d²) hidden space. Theorem 1 gives the gradient step in O(d).
• A second-order method. LIT-535, Theorem 5.1: for any k there exist weights predicting x_testᵀ ŵ_k^Newton from M_j = 2M_{j−1} − M_{j−1} S M_{j−1}, M_0 = αS, S = XᵀX, in k + 8 layers with O(d) hidden dimension.
• Trained models do land in the frame. Across all three papers, transformers trained on in-context regression improve their predictions monotonically with depth and converge to the problem’s optimal estimator. Whatever they are running, they are running something iterative on something model-shaped.

And what the same evidence rules out

No construction of this kind can identify the algorithm. That follows from the constructions themselves: the architecture expresses first-order, closed-form and second-order solvers alike, at comparable and modest cost. A proof that a transformer can implement algorithm A is therefore evidence about the architecture, not about any particular trained model — and it is the step most often skipped when this literature is summarized.

That corollary is the reason behind two methodological practices. It explains why an expressivity claim and an emergence claim have to be tested on different models SOTA-323: a construction or a purpose-trained model shows only what weights exist, and says nothing about what the pretraining objective produces. And it explains why rival accounts of the algorithm must be separated by convergence rate and conditioning rather than by output fit SOTA-324: when the architecture admits gradient descent, ridge and Newton alike and all converge to the same answer, only a property on which they must differ can say which one a trained model runs.

This is why the record files the identification separately, as THEORY-068, and why that document is Rejected while this one is Active. ADR-031: an account can be wrong about the reason while the thing it was invoked to explain goes on happening.

Standing

Active. Nothing in the dispute touches the constructions; LIT-536, which argues hardest against the gradient-descent reading, is explicit that what the constructions establish is architectural expressivity and lets that stand.

What would change this

A demonstration that in-context learning in some model is not organized around anything activation-resident and model-shaped — for instance a mechanistic account of a competitive in-context learner whose per-layer behaviour has no iterative reading at all. LIT-535’s LSTM result is the nearest thing: trained identically, LSTMs weight recent examples more than early ones and their predictions do not improve across layers, which shows this frame is architectural rather than generic. It is a limit on the frame’s scope, not a counterexample within it.