Preprint reports lower compute use when AI models manage their own context

A preprint introduces Context Language Models, which manage the information they retain by editing a context file themselves. The researchers report that applying this approach to existing models produced 11.4% higher accuracy with 21.5% fewer FLOPs, a measure of computation, on BrowseComp-Plus. On 12-hour EdgeBench tasks, they report 5% higher scores with 59% fewer FLOPs. The work also introduces Suffix Cache Reuse for serving these models, reporting 35% less server-side computation than standard SGLang at matched performance. These are preprint results on the tested tasks and serving comparisons.