Topic: rendering

RENDER is a new benchmark introduced to evaluate language models' memory by controlling how reader-facing evidence is presented, addressing limitations in current evaluations that treat input history inconsistently. This matters because it ensures more standardized and fair assessments of memory capabilities across different systems.

arxiv.org↗