- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The authors explore the underlying meaning behind the LLMs’ weight growth during the training as was discovered by the previous works: “learning or merely optimization effort.” In doing so, they build a testing framework based on the WikiText by incorporating a shuffling-repetition scheme. The authors conduct experiments on the grid of the described parameters without altering the optimization process and report the results. The findings reveal that the high weights’ magnitude does not necessarily mean better generalization. In fact, the largest magnitude is achieved by the experiments with the fully shuffled tokens and high repetition. By exploring other parameter combinations, the authors describe three regimes: “LEARNED, MEMORIZED, and NOTHING-TO-LEARN.”
Critical strengths
- The experiments setup with the idea of the two parameters (shuffling-repetition) is a clean way to modify the data’s information capacity to suit the authors’ needs.
Critical weaknesses
- The main problem is the lack of novelty. I do not see why this is in any significant way different from the classic ML overfitting theory. The same results can be achieved with the simple logistic regression: when there is no underlying learning signal (such as fully shuffled data) and the model is forced to minimize cross-entropy (high repetition), the optimizer simply memorizes the noise by constructing an arbitrary separating hyperplane. To maximize the margin on this arbitrary boundary, the model pushes its weight magnitudes to infinity. Therefore, the observation that fully shuffled, highly repeated data produces the largest weight-scale growth is a predictable consequence of margin maximization, not a novel discovery about transformer training. The paper in its current state simply adds experimental complexity to a phenomenon that is already well-understood in classical ML theory.
- The authors note that "transformer weights grow systematically during training (Merrill et al., 2021)" and later cite Kosson et al., 2024. The latter is the more recent work that shows that unconstrained growth is actually not correct. Weight decay balances the weight increase, eventually reaching the "rotational equilibrium". The authors claim that the tracks "accumulated radial-alignment work." However, the summary admits that in high-repetition setting, weight growth shrinks due to WD (as the rotational equilibrium shows). Then, is not a measure of "training effort". Instead, the metric is heavily affected by the AdamW hyperparameters. Using a metric that is actively affected by the optimizer's regularization to measure "how much the optimizer is working" is a flawed premise.
Questions for authors
- 8000 steps is a small budget for the training, where early-training dynamics effects can be still influential. Could you perhaps repeat your experiments while scaling down your model size? The current number of tokens (98M) seems too low to make any conclusions (way too far Chinchilla optimality).
- How important is the learning rate decay to the obtained results? Optimization dynamics say that under a constant learning rate, the opposing forces of gradient updates and weight decay should eventually stabilize at a steady state (rotational equilibrium). When the learning rate decays, this equilibrium boundary artificially drops, allowing weight decay to dominate and pull the weight magnitudes down. What happens to these weight scale trajectories if you run the exact same experiments but keep the learning rate strictly constant?