State space models are known for efficient linear-complexity processing of long sequences, but their performance often degrades as context length increases. The new arXiv preprint introduces SuPER-Mamba, a training-free framework that aims to address this issue.
The abstract explains that SuPER-Mamba uses surprisal—a measure of how unexpected a token is—to prioritize routing in extended contexts. By doing so, it seeks to counter the quality drop that typically accompanies longer inputs, without requiring additional training.
Because the abstract is truncated, details on the exact mechanism and experimental results are not yet available. The paper is posted as arXiv:2610.16636 and has not been peer-reviewed.