Thread Of Thought: Definition and Examples

Thread of Thought (ThoT) is a prompting strategy introduced in 2023 by researchers to improve the ability of large language models to handle large and complex contexts. Unlike Chain of Thought, which breaks down a problem into successive logical steps, Thread of Thought asks the model to "follow the thread" of the available information, extracting and connecting relevant elements before constructing its response.

The principle is based on a simple but powerful instruction: the model is asked to first review the entire provided context, identify the segments of information relevant to the question, and then weave these elements together into a coherent reasoning. This approach is particularly effective when the answer requires synthesizing information scattered across a long document or multiple sources.

The major advantage of Thread of Thought over other reasoning techniques is its ability to handle information noise. In a rich context containing both relevant and irrelevant information, ThoT guides the model to naturally filter what matters, thus reducing hallucinations and off-topic responses.

This technique stands out for its simplicity of implementation: it is usually enough to add an instruction like "Carefully go through the context above, identify relevant information segment by segment, then synthesize your answer" to obtain significant improvements on comprehension and question-answering tasks.