Large reasoning models (LRMs) generate long chain‑of‑thought traces before producing an answer, which incurs substantial inference cost. Pruning can cut this cost, but its success hinges on the calibration data used to gauge parameter importance. Prior work calibrates on the model's own rollouts yet treats every reasoning token uniformly, ignoring whether it actually contributes to a correct solution. Consequently, pruning tends to keep weights that are statistically salient rather than those that drive correct reasoning, allowing erroneous computation paths to survive and degrading both accuracy and trace length. To address this, we introduce OBC‑Prune. OBC first builds difficulty‑matched pairs of correct and incorrect rollouts from problems where the model gives inconsistent answers, then applies intervention‑based analysis to estimate the causal importance of each reasoning sentence—measuring how removing its influence changes subsequent predictions. These causal scores are converted into per‑token weights that rescale the calibration activations used by one‑shot pruning methods such as SparseGPT, Wanda, and ALPS, without altering the underlying pruning algorithms. Experiments on DeepSeek‑R1‑Distill‑Qwen 1.5B, 7B, and 14B models at 40% and 50% sparsity show consistent gains over state‑of‑the‑art calibration baselines across most model sizes and sparsity levels on MATH500, LiveCodeBench, and AIME 2025. The results demonstrate that preserving causally important reasoning circuits is far more effective than uniformly preserving observed activations.
Review