Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kStrict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B)–Accuracy: 93.69
mathematical-reasoning-on-aime-2025Strict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B)–Accuracy: 35.1
mathematical-reasoning-on-aime-2026Strict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B)–Accuracy: 38.44
mathematical-reasoning-on-aime24Strict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B)–Acc: 43.02
mathematical-reasoning-on-minerva-mathStrict top-16 (Qwen3-235B-A22B-Instruct-2507 -> Granite-4.1-8B)–Accuracy: 44.11