I love the silliness of the effect and how, like many years past, you have to resize to discover it. This post about it on Bluesky warms my heart:
“这张证书是正大集团在中国的起点。”泰国正大集团资深副董事长谢毅介绍，40多年来，正大集团在中国设立企业670家，年营业额突破2000亿元人民币。作为改革开放后率先进入中国的外企之一，正大集团见证了中国从打开大门谋发展到成为世界第二大经济体的过程。
。关于这个话题，搜狗输入法提供了深入分析
Surprisingly, I also found that despite the training reward being significantly higher, “best-of-N” distillation underperforms both CISPO and MCTS on the eval suite. While it’s not entirely clear why, we can theorise: if our model has a 98% chance of making at least one reasoning error during its thinking trace, there’s still a $1 - 0.98^{64} \approx 72.6 \%$ chance of selecting at least one correct trajectory. But if there’s no incentive to produce robust reasoning every time, it’s unlikely the model will learn to develop strategies that improve its single-shot score. In secondary school I used a number of techniques to keep track of intermediate steps when solving maths problems. This significantly reduced the probability of making “dumb mistakes” in exams. If I had the option to take the exam multiple times I would never have adopted those techniques!
Avian has slowly but surely been reaching closer parity with other big physics engines