← All questions
LLM FoundationsHard· Pre-training· Attention· Bias

If primacy comes from the system message always sitting...

The system-message explanation only accounts for the instruction-tuned phase of model development. Base models, however, exhibit a related bias because of their pre-training objectives.

In large-scale pre-training, documents often front-load context, making the initial text of a document disproportionately predictive of the subsequent content. This creates a pre-training-level prior for primacy. When a model is later instruction-tuned with a fixed system message, these two priors compound, reinforcing the U-shaped performance curve.

Code playground (Python)