Get the latest tech news

Self-reasoning tokens: teaching models to think ahead

What is the mathematical formulation of reasoning? How can we make LLMs like chatGPT think before they speak? And how can we make that baked into the model so it can learn to think in a self-supervised way without having to "explain it step by step" (or another famous prompt

the authors carefully crafted a mathematical formulation to impede what they call "Pre-Caching," or the ability of the model to make intermediary computations that would be useful beyond the very next token. As a sanity check, this gap should increase as the length of the predicted text grows because the model would have more tokens to produce said computations, and indeed, that was what they found in the paper. The experiment also shows that the model benefits from having multiple tokens to do its "reasoning," forecasting the capability to form long-range dependencies.

Get the Android app