[Objective] To elucidate the mechanism by which seasonal freeze–thaw cycles, coupled with initial water content, attenuate the collapsibility of Ili loess; to quantify the linkage between microstructural reconfiguration and macroscopic collapse deformation; and to provide a theoretical basis for landslide prevention in seasonally frozen loess regions.[Methods] Freeze–thaw cycling tests were performed with varying initial water contents and numbers of cycles, in conjunction with laboratory collapsibility tests and scanning electron microscopy (SEM). Image- and statistics-based analyses quantitatively characterized pore-size distribution, orientation frequency, pore abundance, and pore fractal dimension. Correlations with the collapsibility coefficient were established to form an integrated macro–micro analytical framework.[Results] (1) With increasing freeze–thaw cycles, the collapsibility coefficient decreased significantly for all water contents, with low-water-content soils being more sensitive. For a water content of 14.2%, the peak collapsibility coefficient decreased from 0.094 to 0.084 after nine cycles (≈10.6% reduction); for 20.2%, it decreased from 0.079 to 0.076 (≈3.8% reduction). (2) Freeze–thaw weakens interparticle cementation and alters intergranular contacts and fabric, driving internal structural reorganization toward a new stable configuration. (3) The pore system evolves systematically: the proportion of large pores declines, the pore-size spectrum becomes finer and more uniform, orientations are re-arranged, and both pore abundance and pore fractal dimension change accordingly. (4) Microstructural statistics are consistent with qualitative observations and the macroscopic evolution of collapsibility.[Conclusions] Freeze–thaw cycles reduce interparticle bonding and remodel the pore–particle fabric; the diminution of large pores together with changes in fractal dimension constitutes the core mechanism governing the macroscopic weakening of collapsibility. On this basis, a physical process model—“freeze–thaw → microstructural reconfiguration → collapsibility response”—is established, providing theoretical support for the monitoring and mitigation of loess landslides in seasonally frozen regions such as the Ili area.