The paper proposes a framework for computational coarse-graining, where a sequence of updates is replaced by a cheaper module that still preserves task performance. The authors argue that evaluation must be continuation-relative: a replacement should be judged not only on its direct output, but on how that output influences subsequent computation in a recursive expert system.
By examining interfaces between the original and replacement modules, the framework aims to predict when a coarse-grained approximation will fail. The abstract highlights failure modes that arise when a replacement disrupts the downstream computation, even if it appears accurate in isolation.
This is a single preprint from arXiv, so no conflicting sources are available. The full details of the framework and its experimental results are not included in the abstract, but the core idea is clear: cheap replacements must be validated in context, not in isolation.