Beyond Policy Rollouts: Embedding-Space Directional Exploration with Anchored Noise for LLM RLVR

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kEDEN (Qwen2.5-Math-7B)–Accuracy: 90.3
math-word-problem-solving-on-mathEDEN (Qwen2.5-Math-7B)–Accuracy: 81.8
mathematical-reasoning-on-aime24EDEN (T=1.4, Qwen2.5-Math-1.5B)–Acc: 20.0
mathematical-reasoning-on-minerva-mathEDEN (Qwen2.5-Math-7B)–Accuracy: 36.03