Terrestrial palaeo-temperature data are of great value in improving our understanding of past climate and they provide a basis for evaluating climate simulations. Such data are, however, poorly constrained for long time-scales. In addition to the scarcity of high-quality continuous time-series, finding proxies with a clear response to past temperature changes and developing appropriate reconstruction methods are major challenges. We present a new and robust method – Locally-weighted Weighted-average partial least squares (LW-WAPLS) to reconstruct quantitative temperature changes based on a high-resolution 1.74-Ma pollen record from the Zoige Basin on the eastern Tibetan Plateau, where the vegetation today is mainly controlled by temperature. The reconstructed mean annual (MAT) and warmest month (MTWM) temperatures reveal a general cooling trend with two major shifts at ~1.54 and 0.62 Ma BP, and regular glacial-interglacial variability ranging from ~ − 4 to 2 °C and from 8 to 16 °C, respectively. They indicate ~4–5 °C (MAT) and ~ 5–6 °C (MTWM) magnitudes of glacial-interglacial temperatures. Both statistical and ecological evaluations validate the reliability of the reconstructions. The reconstructions provide important insights into the spatial aspects of long-term terrestrial temperature change. LW-WAPLS shows advantages over both the traditional modern analogue technique and non-linear transfer-function methodologies such as WAPLS for reconstructing the broad-scale climate changes for the Zoige Basin, by combining the strength of both methods. The LW-WAPLS approach potentially provides a robust tool to develop pollen-based climate reconstructions over long time-scales.