ZeroSearch model outputs malformed format

[YiboXue]

I’m running ZeroSearch and noticed the model output format is often malformed, even on simple questions. Here is a representative example output:

<think> when did the president who set the precedent of a two term limit leave office </think>
<information>
<answer> 1989 </answer>

why this happen

[sunhaonlp]

Hi, thank you for your interest in our work! During training, we did not incorporate a format reward, which may lead to malformed responses in certain cases. Adding a format reward during RL training should help address this issue.

[YiboXue]

but 3b output like this, by using inference.py in the project