Validate the attention-dilution mechanism underlying batching regularization

Confirm rigorously whether the conservative editing shift produced by batching independent sentences in prompted large language models is caused by self-attention dilution and the resulting reduction in attention budget per token.

Background

The paper finds that increasing the number of independently submitted sentences packed into one prompt generally reduces word-level edit rates and overcorrection in minimal-edit grammatical error correction. This effect improves precision but can reduce recall and fluency, suggesting that batching functions as a behavioral regularizer.

The proposed explanation is that the self-attention softmax normalization distributes a limited attention budget across more sentences, weakening the model’s tendency to perform stylistic rewrites and causing it to prioritize salient grammatical errors. The paper explicitly characterizes this explanation as an empirical conjecture and notes that rigorous confirmation would require controlled probing, particularly on open-weight models.