THEORY-073: Training passes through stages marked by changes in loss-landscape degeneracy, and the loss curve does not show them

Source

Hoogland et al. (2024), LIT-543 — ARXIV-2402.02364. The method — estimate the local learning coefficient through training and take critical points of its curve as stage boundaries — and both case studies below are this paper’s.

The account

λ is THEORY-075’s exponent: that account establishes it as the right measure of effective complexity, as a statement about the Bayesian posterior, and this one tracks it through a training run.

If λ measures how degenerate the loss landscape is around the current parameters, then watching λ through training watches something the loss does not report: not how well the model fits, but how much room it has to vary without changing what it computes. The claim is that this quantity moves in distinguishable periods, that the boundaries between them are real, and that they coincide with the formation and reorganisation of internal structure.

The method inverts the usual order. A progress measure is built from a hypothesis about the mechanism and then tracked. Here the boundaries are found first, from one setting-agnostic curve, and interpreted second — which is why the paper calls degeneracy an “unsupervised” alternative, and why it is honest that “once a change is detected through its effect on degeneracy, it remains to interpret the change”.

What was measured

Two-layer attention-only language transformer. Five stages ending at t = 900 / 6.5k / 8.5k / 17k / 50k, with Δλ̂ = +26.4 / +22.5 / −1.57 / +8.62 / +1.77 against Δℓ̂ = −2.33 / −1.22 / −0.18 / −0.40 / −0.34. Bigram statistics, then common n-grams, then previous-token heads, then the induction circuit of Olsson et al. A single training run recapitulates the progression Olsson et al. found across fully-developed models of increasing depth.

In-context linear regression transformer. Five stages ending at t = 1k / 40k / 126k / 320k / 500k, with Δλ̂ = +21.4 / +149 / −12.3 / −44.1 / +3.56. The model learns the optimal context-independent prediction, then acquires in-context learning — and then loses it across LR3 and LR4 while specializing to its pre-training task distribution, with layer-normalization weights collapsing to zero, all while the loss keeps falling.

In a toy model, the same structure is derived rather than observed. LIT-544 shows that in the Toy Model of Superposition the critical points governing the posterior’s phases are the same ones that explain SGD’s plateaus — which is the closest thing to a mechanism this account has.

Why it is Proposed

The evidence is coincidence, and the authors say so. “Suggestive evidence that degeneracy and development are linked.” Stage boundaries land near structural changes in two models; nothing derives one from the other, and with five boundaries per model the argument is qualitative.

Two case studies, labelled as two case studies. “We do not claim that the structural and behavioral developments we observed in each setting are universal phenomena.” Neither model is large, and for both of them the final stage had no change the authors could find.

The stage rule is a convention. “Critical points of the LLC curve” is reasonable and is a choice; the estimator is an SGLD sample with a localizing radius, so the curve’s fine structure is partly the sampler’s.

And the decreases are unexplained. λ falls in some stages, in both models. Saddle-to-saddle theory predicts complexity increasing; the transitions sketched in the source’s own §5 are increases. Lowering λ at constant loss also lowers the free energy, so it is permitted — but, in the authors’ words, “providing a full theoretical account of these stages is an open problem”. An account of development that cannot say why development sometimes runs backwards is incomplete on its own terms.

What it is up against, and what it is not

It is a rival to LIT-085’s progress measures and the source names it as one. SOTA-200 records that those measures are computed by projecting onto five frequencies the network was reverse-engineered to be using, and are meaningless without that reverse-engineering. This needs no such prerequisite — and returns a boundary rather than a mechanism, which is less than the reverse-engineering gives.

It is not a claim that λ explains the stages. Only that it finds them. Whether the degeneracy change causes the structural change, is caused by it, or is the same event under two descriptions, is not addressed.