A simple trick to change how LLMs reason: every so often, inject "Wait, but " into the in-progress reasoning token stream.
- A simple trick to change how LLMs reason: every so often, inject "Wait, but " into the in-progress reasoning token stream[adf].
- This forces the LLM to reflect on what it might have missed, giving more resilient output.
- Another hint that managing contexts will be the "montage moment" for LLMs.