From Response to Query: The Power of Reverse Thinking in Language Models

In a new paper Time-Reversal Provides Unsupervised Feedback to LLMs, a research team from Google DeepMind and Indian Institute of Science proposes Time Reversed Language Models (TRLMs), a framework that allows LLMs to reason in reverse—scoring and generating content in a manner opposite to the traditional forward approach.