Research digest The research that actually changes what you build. arXiv, lab papers and reproducibility notes

inference performance

Research digest

Length-Bucketed Batching: How to Stop Wasting GPU Cycles on Padding for Small Language Models
Byeditor September 26, 2026September 26, 2026

The padding problem: why naive batching wastes compute Processing one ticket per forward pass is described in the article as “the single largest source of waste in the whole pipeline. ” The reason is memory bandwidth.

Read More Length-Bucketed Batching: How to Stop Wasting GPU Cycles on Padding for Small Language Models

AI research and scientific papers translated for working engineers: the method before the result, the baseline flagged when it is weak, and a link to the artifact every time.

The research that actually changes what you build.