nabla-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space

Benchmark Model Rank Results
mathematical-reasoning-on-math-500โˆ‡-Reasoner (Qwen-2.5-7B-Instruct)#91Accuracy: 80.4
mathematical-reasoning-on-aime24โˆ‡-Reasoner (Qwen-2.5-7B-Instruct)#107Acc: 26.6
mathematical-reasoning-on-aime-2025โˆ‡-Reasoner (Qwen-2.5-7B-Instruct)#127Accuracy: 20.0