This article explores a surprising and increasingly well-supported idea: words like “Hmm,” “Wait,” or “I cannot” in reasoning models are not just stylistic filler, but powerful control signals. It shows how these seemingly ordinary tokens act as mode switches that influence how deeply a model thinks, whether it backtracks, and even how it behaves under safety constraints. Drawing on recent research, the post explains why suppressing these tokens degrades reasoning performance, while forcing them can sometimes lead to better answers.

The article introduces the idea of tokens as a form of self-conditioning, where each generated word reshapes the model’s internal state and future trajectory. It dives into mutual information spikes and reveals how “thinking words” often align with moments of sudden informational gain inside the model. Practical techniques like budget forcing and test-time compute scaling are examined through this lens, reframing them as ways of speaking to the model in its own language.

Going beyond reasoning tasks, the post shows how similar token-level switches appear in safety alignment, where the first few words can determine whether a model refuses or complies. It proposes a practical taxonomy of “special tokens,” including reflection markers, commitment signals, and critical failure points. The conclusion argues that chains of thought are best understood not as transcripts of reasoning, but as handles for steering computation.

👉 Read the article here

‍