Human behaviour is uniquely flexible: we can associate almost any stimulus with any action and readily change these associations based on the current goals and rules of our task. This ability is widely thought to depend on frontal and parietal cortex. In particular, a specific network of regions, situated mainly in lateral prefrontal cortex (PFC), anterior insula, dorsomedial frontal cortex, and intraparietal sulcus (Figure 1a), is active for many types of cognitive demand 1, 2, 3 and linearly predicts deficits in fluid intelligence after focal lesion [4]. This network is commonly referred to as the multiple-demand (MD) network [5], frontoparietal control network [6], or cognitive control network/networks 7, 8.

A key proposal for the function of the MD network is that it integrates many different types of information from distributed specialised processing systems into a useful structure according to the roles and relationships required by particular tasks [9]. This is appealing because it provides a basis for creating the bespoke arbitrary associations between stimuli and responses that are needed for flexible goal-directed behaviour [10]. For example, in computational models where PFC neurons are randomly connected to sensory neurons, PFC responses represent arbitrary combinations of stimuli 11, 12. If the strengths of these connections also update rapidly, for example, through Hebbian plasticity [11], and if the framework extends not only to fixed sensory neurons but also to motor programmes and cells in other domains, the result is a theoretical basis for how almost any arbitrary association can quickly be encoded and acted on (Figure 1b). In line with these proposals, nonhuman primate PFC neurons encode a wide variety of information types, including combinations of stimuli and tasks [13], and exhibit rapid history-dependent changes in selectivity [14]. Moreover, human frontoparietal cortex shows highly flexible connectivity patterns, which adjust across tasks 15, 16, and both large-scale functional connectivity [17] and individual differences in integration among frontal and parietal regions [18] predict cognition. Exciting recent work further specifies that the extent of a region’s cross-network connectivity predicts the number of different task elements that can be decoded from it, with the specific profile of information coding in each region predictable from the information held in connected brain regions [19].

A broadly connected, rapidly updating, general processing associative resource in frontoparietal cortex could therefore provide a way for the brain to flexibly route a wide variety of stimulus and task information to responses as needed for goal-directed behaviour. Two hallmark predictions for such a radically flexible system are that it should (1) represent a wide variety of different information across tasks and (2) demonstrate adaptive tuning according to current attentional focus. A key manifestation of flexibility is therefore selective (but not fixed) prioritisation of attended information. A radically flexible system may also adjust the relative selectivity of its coding scheme according to wider task demands. Furthermore, these changes should happen on a timescale suitable to underpin the momentary focus of real tasks, should have downstream consequence on information processing elsewhere in the brain, and should be readily predictive of human behaviour.

A powerful tool to examine these predictions is multivariate pattern analysis (MVPA) for neuroimaging, which looks for patterns of activity that reliably distinguish between conditions or are organised with predictable geometry, to quantify representational content 20, 21. Here, we review the recent human MVPA literature and show that neural representations in frontoparietal cortex indeed exhibit highly flexible information coding in line with the above predictions. However, the mere observation of decodable information is not sufficient to conclude that we have captured aspects of neural activity that are critical for behaviour. Instead, we need to draw out the decoding–behaviour association explicitly, seeking to understand when and how information coding is and is not tightly predictive of behaviour. The importance and challenge of linking neural representations to behaviour is increasingly reaching consciousness for the field; we review recent attempts to make this link, showing that some — but not all — decodable information can be directly linked to human behaviour.