The human brain quickly adapts to a changing environment, often faster than can be explained by synaptic plasticity. One of the main candidates to account for such rapid changes is neuromodulation. We explore how modulation can assist a recurrent neural network to accommodate a range of variations of the tasks it is initially trained on. There are a number of approaches allowing for rapid acquisition of new tasks, for example, the neuronal gain modulation model [1]. However, we found that by affecting the input-output function of a neuron such an approach can neither address the variation of a simple pattern discrimination task nor accommodate the simultaneous impact of different neurotransmitters. Here, we investigate a more general model of neuromodulation in which we can adjust gains at the level of synapses [2]. Modulation is achieved through element-wise multiplication of the circuit connectivity with a modulation matrix, obtained as a linear combination of fixed synaptic masks. The masks represent the susceptibilities of the individual circuit synapses to neurotransmitter release from a given modulatory neuron. Modulatory firing rates thus serve as coefficients of the linear combination, and are presumably provided by an external, contextual input (Fig. 1). We compare neuronal and synapse-specific modulation in two different tasks, i.e. motor control and pattern discrimination (Fig. 1.2). We observe that an appropriate combination of just a few synaptic masks can effectively change network connectivity (Fig. 1.4). As expected, more intricate, higher-dimensional variations require more components in the linear combination composing the modulatory mask. To investigate this interdependence between the (minimum) number of modulatory controls and the variation dimensionality, we perform SVD decomposition of the stacked modulatory matrices that are analytical solutions [3] for different variations of the pattern discrimination task (Fig. 1.3). We observe that the distribution of the singular values reflects the amount of independent modulatory controls required to address the variation. Our approach provides a useful framework for improving the mechanistic understanding of rapid adaptation in neural circuits. It shifts the focus from synaptic plasticity as a learning mechanism for one task to flexible neuromodulation that fits an entire subspace of tasks.