Decoding Tree Sketching (DTS) was proposed in November 2025 by a research team from Johns Hopkins University, the University of North Carolina at Charlotte, and other universities. The related research results were published in the paper "...".DTS: Enhancing Large Reasoning Models via Decoding Tree Sketching".

DTS is a model-agnostic decoding framework that mitigates overthinking by constructing a dynamic inference tree during inference. Specifically, DTS uses next-label entropy to locate critical divergences, guiding the selective expansion of branches that form the backbone of the inference tree. All branches are generated through autoregressive parallel decoding, and DTS employs an early stopping strategy to return the first completed branch, which corresponds to the shortest inference path and serves as an approximation of the shortest optimal trajectory.