On the Edge of Understanding: Churn in LLM Post-Training

Beneath steady accuracy improvements in LLM post-training, prior work finds that many queries flip between solved and unsolved across training epochs. However, the underlying mechanisms driving these flips remain poorly understood. Through large scale analysis, we find a single gradient update flips $21\%$ of learnable queries between solved and unsolved throughout training, across both in-distribution and out-of-distribution datasets — a phenomenon we term $\text{LLM churn}$. Accuracy gain is the surplus of positive over negative churn, both persisting even after accuracy plateaus. Churning queries reside on LLM's edge of understanding, razor-thin confidence margin between correct and incorrect answers makes correct answer $12.5\times$ more vulnerable to perturbation. Persistent churn incurs substantial cost: solved queries flips to unsolved lead to $25.8$ pp forgetting gap, and $28.5\%$ of final accuracy comes from unreliable answer on queries that flips repeatedly across training checkpoints. Finally, we find that churn is not benign: applying sharpness-aware minimization after accuracy plateaus reduces churn while lifting final accuracy by $5$ percentage points for models up to $4B$ parameters, with no benefits on larger models. Our work highlights the importance of analyzing the evolution of query-level performance during post-training and points to a promising path for improving learning efficiency via reduced churn.