Occupancy predictors typically rely on large, persistent banks of Gaussians or learned queries whose size is fixed by hand. This makes pruning attractive, but the usual shortcuts—opacity, activation frequency, or parameter magnitude—are poor guides. According to the paper, these proxies do not measure whether removing a component actually degrades the prediction.
Decision-FLIP addresses this by making the pruning decision depend on the component's measured effect on the model's output. Instead of ranking components by a heuristic, it directly evaluates the consequence of removal. The method is also budget-free, meaning it does not require a predetermined target size; the model itself determines how many components can be safely dropped.
The significance is practical: if the approach holds up, it could let occupancy models run more efficiently without the laborious tuning of pruning budgets or the risk of silently removing important structure. As a single arXiv preprint, the claims have not yet been peer-reviewed or independently replicated, but the core critique of existing proxies is well targeted.