interpretabilityfactneutralSteering away the '什么意思' meta-token makes the model answer ambiguous text literally instead of catching wordplay or punsAI Alignment Forum28 Jul 2026https://www.alignmentforum.org/posts/6ek6n7yZ5DzfarJHy/towards-surfacing-model-algorithms-with-meta-tokens-in-the-j