A Reproducibility Study of Partial Residual Ablations in Pre-LN Transformers

Modern artificial intelligence that writes, translates, and reasons relies on a specific architectural design known as the transformer. At the heart of this design are deep stacks of processing layers, each one refining the information it receives from the previous one. To keep these deep networks from falling apart during training, engineers use "skip connections." Imagine a river flowing through a series of waterfalls; the skip connection is like a bypass channel that allows water to flow around a waterfall and rejoin the stream downstream. This ensures that the signal never gets lost or diluted, no matter how many layers the data must pass through. For years, the standard practice has been to include these bypasses after every single processing step. But a recent investigation asked a simple, structural question: what happens if we remove just one of these bypasses? Does the system fail completely, or does it find a way to adapt?

A researcher named Pratik Kumar Babariya set out to answer this by building artificial intelligence models and systematically removing these safety nets. He tested two specific scenarios: one where the bypass after the attention mechanism was removed, and another where the bypass after the feed-forward network was removed. The attention mechanism is the part of the model that decides which words in a sentence are most relevant to each other, while the feed-forward network is the part that processes and transforms that information. The goal was to see if the model could survive without these specific shortcuts, or if the removal would cause the entire learning process to collapse.

The initial experiments, conducted on smaller models, suggested a bleak outcome. When the researcher removed either of the two bypasses, the models failed to learn, getting stuck at the same poor performance level as a model that had no bypasses at all. It appeared that both shortcuts were equally essential. However, this early result was likely a trick of the measurement tools. The researcher realized that the way he was evaluating the models was introducing random noise, making it impossible to see subtle differences in performance. He rebuilt the experiment with a more rigorous setup, using fixed data for testing and ensuring that every variable was controlled.

When the experiment was run again with this cleaner method, a striking asymmetry emerged. Removing the bypass after the feed-forward network caused the model to fail completely, just as the initial tests had shown. The model could not learn, and its performance remained stuck at a high error rate. But when the bypass after the attention mechanism was removed, the model did something unexpected: it learned. It did not perform as well as a model with all its shortcuts intact, but it successfully learned to process the language, achieving a level of performance far better than the failed model. This result was not a fluke; the researcher ran the successful configuration eight times with different random starting points, and in every single case, the model learned effectively. The failure of the other configuration was equally consistent, happening every time it was tested.

The researcher then looked at why this difference existed. He observed that in the failing models, the flow of information from the bottom layers to the top had completely stopped, a phenomenon known as gradient starvation. This happened in both failing cases. However, in the model that succeeded, the attention mechanism seemed to find a workaround. Because the attention mechanism looks at all the words in a sentence simultaneously, it can learn to route information across different positions, effectively rebuilding a path for the data to flow even without the physical shortcut. The feed-forward network, by contrast, processes each word in isolation and cannot create this kind of cross-connection. Without its bypass, it has no way to recover the lost information.

The story does not end with a perfect, solved mystery. When the researcher scaled the experiment up to a much larger model, the results became less clear. While the model that removed the feed-forward bypass still failed, the model that removed the attention bypass showed signs of learning, but its performance varied wildly depending on the random starting conditions. In some runs, it learned well; in others, it struggled. This suggests that while the ability to adapt is real, it becomes harder to maintain as the model grows larger and more complex. The researcher also uncovered a significant methodological error in his own early work: he had tried to fix the failing models by adjusting a multiplier during the training process, which made them look successful. He later realized that the training algorithm had automatically compensated for this change, making the adjustment an illusion. By correcting this mistake and reporting the failure transparently, he ensured that the final conclusion was based on reality, not an artifact of the software.

The core finding is that not all shortcuts in an artificial intelligence brain are created equal. Removing the path after the part of the model that connects words to each other allows the system to survive and learn, likely because that part of the system can rewire itself to compensate. Removing the path after the part that processes individual words causes the system to break down completely. This distinction holds true for smaller models and is strongly suggested for larger ones, though the larger models require further study to confirm the limits of this adaptability. The work serves as a reminder that in complex systems, the specific arrangement of connections matters as much as the connections themselves, and that even when a system seems to fail, a closer look can reveal a hidden capacity to recover.