Abstract
Learning to avoid harm is essential for wellbeing, and its dysregulation is thought to underlie anxiety-related disorders. To date, there is no generally accepted computational theory of human avoidance learning. A potential reason is the absence of systematic empirical benchmarks capable of distinguishing competing theories. Here, we test 12 canonical and novel associative learning phenomena in a naturalistic avoidance paradigm across six preregistered virtual reality experiments in healthy humans (total N = 599). Our findings converge on three computational principles. First, avoidance learning is predominantly inference-based: participants updated avoidance following signalled US omission and transition structure changes. Second, it reflects cost-benefit integration: spontaneous extinction and retrospective revaluation emerged under naturalistic motor costs, but gave way to residual avoidance when costs were removed. Third, learning about compound cues suggests configural rather than linear feature integration. These principles, and the detailed pattern of findings, can guide the development of candidate computational theories.