To investigate how deep-water hypoxia induced by seasonal thermal stratification affects fish distribution patterns in both vertical and horizontal dimensions, hydroacoustic surveys and gillnet sampling were conducted in Niutoushan Reservoir in 2025, combined with vertical profiles of water temperature and dissolved oxygen. The compression of vertical fish habitat and changes in spatial aggregation patterns were analyzed. To address the difficulty of distinguishing whether the upward shift of fish distribution resulted from hypoxia avoidance using conventional correlation analyses, null models were constructed to test whether fish utilization of hypoxic layers was lower than random expectations. Meanwhile, horizontal fish aggregation hotspots were transformed into station-scale aggregation intensity metrics, and multiple linear regression models (LMs) were used to examine their relationships with water-depth structure and hypoxia intensity.The results showed that vertical fish allocation patterns differed significantly among the four survey months (PERMANOVA, p = 0.001, R² = 0.277). During the stratified period (May and October), the vertical centroid depth (Zc) of fish distribution became significantly shallower and less dispersed (median: 7.54 m), compared with the mixed period (12.19 m), corresponding to an upward shift of 4.65 m (ΔZc = ?4.65 m). Null-model analyses indicated that fish utilization of deep hypoxic layers was generally lower than random expectations under the 3 mg/L dissolved oxygen threshold. Horizontally, fish aggregation hotspots repeatedly occurred in the river–reservoir transition zone (site S3) across multiple months. The LM results revealed an inverted U-shaped relationship between fish aggregation intensity and water depth, with an estimated optimal depth (D*) of approximately 14.4 m. After accounting for water-depth structure and seasonal differences, oxygen-stress deficit (DOdef) showed a negative association with fish aggregation intensity (β = ?1.151 ± 0.444, p = 0.022).This study establishes an analytical framework integrating null-model testing and aggregation-intensity quantification, providing directional evidence for identifying fish hypoxia avoidance and spatial aggregation responses. The framework may support habitat monitoring and hypoxia-risk assessment for fish communities in similar medium- to deep-water valley-type reservoirs during stratified periods.