Three experiments demonstrated that, following the extinction of an established conditioned stimulus (CS; e.g., tone), the pairing of an orthogonal stimulus from another modality (e.g., light) with the unconditioned stimulus (US) results in strong recovery of responding to the extinguished CS. This recovery occurred to about an equal degree regardless of whether or not initial training contained unambiguous stimulus-reinforcer relationships--that is, consistent CS-US pairings--or some degree of ambiguity, including intramodal discrimination training, partial reinforcement, or even cross-modal discrimination training (tone vs. light). Experiments 1 and 2 demonstrated that this recovery of responding was largely specific to the extinguished CS, but moderate generalization to other stimuli from the same modality did appear. The results are discussed with reference to alternative mechanisms applicable to learning-dependent generalization between otherwise distinct CSs. These models assume that such generalization is mediated by either a shared response, shared reinforcer, shared context, or shared hidden units within a layered neural network. A specific layered network is proposed to explain the present results as well as other types of savings seen previously in conditioning of the rabbit nictitating membrane response.