返回探索

Qwen排列平均与Jev改进对比

作者报告对Qwen的输出进行8种排列平均可将KL误差减少约79%,720种排列减少约97%,而Jev模型的改进相对有限。

If Qwen favors the first option, let every answer take a turn there—then average the probabilities. Averaging 8 permutations reduced its KL error by ~79%; all 720 reduced it by ~97%, relative to average single-order error. Jev improved only slightly.

Qwen排列平均与Jev改进对比 1
· 0 次赞在 X 打开