Weight pruning is a common way to shrink large models, but it often hurts their accuracy. A new arXiv paper asks whether pruning actually damages the reasoning ability of large reasoning models (LRMs) or something else. The authors report that much of the degradation comes from a failure to stop rather than a failure to reason.

In other words, pruned models may still know how to reason correctly but lose the ability to end their response at the right point. This distinction matters because it suggests recovery efforts should focus on restoring stopping behavior, not on retraining reasoning skills.

The paper is a single source, so its claims are not yet independently confirmed. Still, the framing offers a fresh angle on why pruning breaks LRMs and where to look for fixes.