Considerable evidence suggests that decisions are highly sensitive to contexts. However, how experience shapes context-dependent computations remain elusive. In a simple stimulus-outcome association task, subjects learned through experience the probability of reward associated with different visual stimuli. On each trial subjects experienced a single stimulus, but over the course of a block of trials, subjects experienced two stimuli carrying different probabilities of reward. Context was manipulated by pairing stimuli carrying the same probability of reward with stimuli carrying different probabilities of reward in different blocks of trials. In order to measure subjective probability, subjects provided trial-by-trial estimate on the probability of reward associated with the current stimulus. We found that subjective probability associated with a stimulus was affected by the other stimulus present in the context. However, context effect was not universally observed across the entire range of probability – it appeared to be constrained by the variance of experienced outcomes. Stimuli with high variance (50% chance of reward) showed significant context effect, while stimuli with low variance did not (10% and 90% reward). Computational modeling revealed that such effect was reference dependent – the direction of context-induced bias on probability estimates was determined by the variance-weighted difference between local frequency of reward of the stimulus and the average reward frequency of the block (reference point). Using fMRI, we found that the ventromedial prefrontal cortex represents information essential to context-dependent computations for subjective probability. We concluded that variance and reference dependence are two key elements of context-dependent computations for probability acquired through experience.