SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathQwen3.5-35B w/SAGE#8Accuracy: 51.34
math-word-problem-solving-on-mathQwen3.5-35B w/SAGE#10Accuracy: 84.72
mathematical-reasoning-on-olympiadbenchQwen3.5-35B w/SAGE#10Accuracy: 68.26
arithmetic-reasoning-on-gsm8kQwen3.5-35B w/SAGE#32Accuracy: 94.16
mathematical-reasoning-on-aime24Qwen3.5-35B w/SAGE#64Acc: 62.04