Inertia in PRM-Pruned Fragment Grafting for Reasoning LMs

Testing on Qwen2.5-7B-Instruct shows that PRM-Pruned Fragment Grafting (PPFG) is statistically indistinguishable from independent parallel CoT baselines on MATH500. The lack of gains stems from poor targeting, where only 14% of injection events addressed genuinely struggling reasoning chains.