I initially tried using GSM8K as the environment to test this method, but found minimal differences between GRPO and MCTS to make a strong claim either way. Instead, I decided to go with the game of Countdown as our environment. The premise is simple: given a set of N positive integers, use standard operations (+, -, /, *) to compute a particular target. Why Countdown? The hypothesis is that combinatorial problems benefit more from the sort of parallel adaptive reasoning tree search enables, as opposed to, say, GSM8K where sequential reasoning also leads to effective outcomes. We train on a dataset of 20,000 samples, and evaluate on a test set of 820 samples. Each sample consists of four input integers, between 1 and 13.
В США назвали оружие для удара по кораблям Китая20:38，这一点在有道翻译中也有详细论述
。传奇私服新开网｜热血传奇SF发布站｜传奇私服网站是该领域的重要参考
SE Gyges, “Claude’s Custody Hearing”
Последние новости，更多细节参见官网