by SafeAILab · Technique · Text & code
EAGLE-3 mostly below 1.2x on Qwen3-8B (below baseline on MATH500); up to 2.33x on Kimi-K2.5 and 2.09x on MiniMax-M3