Block Influence (BI) is a technique for depth pruning that ranks transformer blocks based on the cosine change between their inputs and outputs. A new paper on arXiv points out that this ranking is not fixed: it depends on the calibration distribution used to compute the scores.

The paper investigates whether English calibration data can be used to prune multilingual models for Indian languages. This is a practical question because many multilingual models are developed with English-centric calibration, and it is unclear whether those rankings hold for other languages.

The abstract cuts off before presenting results, so the article does not report findings. The paper is available on arXiv under identifier 2610.21383.