When auditory and visual cues both indicate a stimulus position, mice integrate these cues linearly (Coen et al., Neuron 2023). This strategy is optimal if the cues are independent and informative. What happens if the reward structure is changed to make this strategy suboptimal? Do mice continue to act additively, and do they learn to use the informative interactions between cues?Mice were trained to indicate the left-or-right position of audiovisual cues — which were presented alone (unisensory trials), simultaneously in the same location (coherent trials) or simultaneously in opposite locations (conflict trials) — by turning a wheel. Initially, conflict trials were rewarded at random. Then, training alternated between auditory-dominant and visual-dominant blocks, in which mice were rewarded for responding to the location of the auditory or visual cue, respectively. Unisensory trials remained rewarded as in the original task. We modelled behaviour using a dynamic Bernoulli generalized linear model (Roy et al., Neuron 2021).Mice adjusted their strategy to respond to a single sensory modality in conflict trials. They continued to act linearly, progressively increasing the weight given to one sensory modality while decreasing the other, despite this strategy reducing performance in unisensory trials.These results suggest that linear combination is the fundamental operation with which the mouse brain integrates spatial multisensory cues. Mice learn by adapting the weights dynamically but do not deviate from this linear combination, suggesting that it is a general heuristic that they adopt even when suboptimal.