Key takeaways
- The premise is an artifact of averaging. Visible discontinuities in a loss curve get studied as conceptual breakthroughs, and the smoothness of the rest gets read as “nothing happening”. The paper argues breakthroughs are frequent throughout training and the scalar loss hides them — the more differently-timed transitions there are underneath, the smoother the sum looks. Smoothness is what many transitions look like when added up.
- Two decompositions, because one is not enough. Disaggregate the aggregate into per-example losses and cluster examples with synchronised changes. But one example can depend on several breakthroughs, and distinct concepts arriving together merge their clusters — so also decompose the loss change along directions in a low-rank subspace of the trajectory.
- POLCA is Loss Change Allocation modified in two ways: an arbitrary orthonormal basis instead of axis-aligned units, and per-example rather than dataset-level attribution.
- The basis is built from curvature. Iteratively, at each checkpoint, project the loss Hessian onto the nullspace of the basis so far and append the top eigenvectors, via Hessian-vector products. The very top eigenvectors reflect local oscillation rather than long-term movement, so directions that do not lower loss over the run are discarded.
- The validation is a negative control. On synthetic arithmetic, clustering the exact loss curves recovers digit positions. Clustering POLCA curves recovers digit positions and the skill of carrying — a concept the aggregate cannot see. Maximum carry fraction for exact-loss clustering is 0.514, which is chance.
- Offered as a route to unsupervised interpretability: find the transitions first, then ask what they are.
- Small scale by necessity — a 9M-parameter 3-layer transformer, POLCA computed every 20 iterations, 50 basis vectors.
Standing in the anthology
It generalises a finding the record filed two days ago. THEORY-028 says one specific plateau is the attention recall circuit being built, shown by patching the circuit in. This says the same kind of thing is happening throughout training and is invisible for a different reason — not because the loss is flat, but because the loss is an average over data and directions that are each doing something abrupt.
Together they make a claim the record can now state: the aggregate loss curve is a lossy projection of training, and it is lossy in at least three identified ways — it time-averages oscillation (LIT-453), it sums over differently-timed transitions (this), and it reads flat while the weights are still moving (LIT-454).
It supplies an instrument the record’s evaluation material lacks. SOTA-198 measures the gradient noise scale from the run in front of you; SOTA-194 asks what a benchmark’s composition is. Nothing in the record says how to find out what a training run is doing while it runs, and the answer “watch the loss” is the one this paper is about.
The cost is real and should be stated: POLCA needs checkpoints, Hessian-vector products at each, and a per-example forward pass at every measurement interval. At 9M parameters that is affordable and the paper says the model size was chosen for exactly that reason.