Insight denotes a mental restructuring that leads to a sudden gain of explicit knowledge allowing qualitatively changed behaviour1,2. Anecdotal reports on scientific discovery suggest that pivotal insights can be gained through sleep3. Sleep consolidates recent memories4,5,6 and, concomitantly, could allow insight by changing their representational structure. Here we show a facilitating role of sleep in a process of insight. Subjects performed a cognitive task requiring the learning of stimulus–response sequences, in which they improved gradually by increasing response speed across task blocks. However, they could also improve abruptly after gaining insight into a hidden abstract rule underlying all sequences. Initial training establishing a task representation was followed by 8 h of nocturnal sleep, nocturnal wakefulness, or daytime wakefulness. At subsequent retesting, more than twice as many subjects gained insight into the hidden rule after sleep as after wakefulness, regardless of time of day. Sleep did not enhance insight in the absence of initial training. A characteristic antecedent of sleep-related insight was revealed in a slowing of reaction times across sleep. We conclude that sleep, by restructuring new memory representations, facilitates extraction of explicit knowledge and insightful behaviour.
The idea that sleep can trigger the gain of insight is associated with the names of famous scientific discoverers3. For example, Nobel prize winner Loewi reported that he woke up with the essential idea for an experimental confirmation of his theory of chemical neurotransmission. Mendeleyev, who laid out the periodic table of chemical elements, reported that his understanding of the critical rule underlying it emerged out of a dream following unsuccessful puzzling with the symbols of the elements. Recent studies in animals and humans provided evidence for the concept that neuronal representations of task stimuli and responses acquired during wakefulness become reactivated during subsequent sleep7,8,9,10. This reprocessing of representations is considered to underlie the consolidating effect of sleep on memory4,11,12,13,14,15, but could also be accompanied by restructuring these representations in memory to enable insight. Here, we tested whether sleep influences task representations in memory such that the gain of insight is facilitated. This was expressed in the extraction of explicit knowledge of a hidden abstract rule in stimulus–response sequences learned previously under implicit conditions.
To grasp the inherently unpredictable phenomenon of insight experimentally, we used a modified version of the Number Reduction Task (NRT; Fig. 1a) originally developed by Thurstone and Thurstone16,17,18. Based on continuous monitoring of subjects' behavioural responses, the task allows the exact determination of the time point when insight occurs, that is, when explicit knowledge of a hidden abstract rule is gained, leading to an abrupt, qualitative shift in responding. On each trial of the task, subjects were asked to transform a given string of eight digits into a new string through a stepwise digit-by-digit application of two simple rules to reach a specific digit indicating the final solution to this string. With increasing practice, the subject's responses become gradually faster on this task. Most important, however, a hidden rule was implemented in the digit strings, which was not mentioned to the subjects and was therefore initially processed at an implicit level without awareness. The time point when a subject gained insight into this rule could be determined precisely because at this time he/she would begin to cut short sequential responding to confirm the final solution in advance. All subjects were first trained on three task blocks to induce mental representations of the task that still remained implicit with regard to the hidden rule during this period. The training period was then followed by an 8-h interval of (1) nocturnal sleep, (2) nocturnal wakefulness, or (3) daytime wakefulness (Fig. 1b). Subsequently, subjects were retested on ten blocks.
Sleep more than doubled the probability of gaining insight into the hidden rule compared to wakefulness. In the sleep group, thirteen out of 22 subjects (59.1%) gained insight at retesting, compared to five subjects (22.7%) in either wake group (χ2 = 8.54, degrees of freedom (d.f.) = 2, P = 0.014; Fig. 2). For subjects gaining insight, the time point of its occurrence (number of blocks after beginning of retesting) did not differ significantly between groups (sleep, 4.5 ± 0.8 (mean ± s.e.m.); wake-night, 6.8 ± 1.5; wake-day, 6.0 ± 1.0; P > 0.28). The two nocturnal and daytime wake conditions excluded the possibility that inferior performance after wake intervals resulted from sleep deprivation or variations in circadian rhythm. Subjective ratings obtained at the beginning of retesting revealed that, as expected, subjects in the sleep condition were less tired compared to subjects tested following a wakeful night (five-point scale: 2.95 ± 0.22 versus 3.64 ± 0.25, P < 0.05), but not compared to those tested after a period of daytime wakefulness (2.41 ± 0.22, P > 0.11). This pattern excludes that the inferior performance in the wake conditions resulted from unspecific effects of tiredness.
To ensure that the influence of sleep was on memory representations established during training before sleep rather than a proactive influence on performance at retesting, in supplementary experiments subjects were tested on 13 continuous blocks either in the morning after sleep (7:00 h) or in the evening after daytime wakefulness (19:00 h), with only ten blocks included in the analysis for the direct comparison with the retest situation of the main experiment. Since, unlike in the main experiment, subjects had no training before these test periods, off-line restructuring of initially acquired task representations could not occur during the periods of sleep or wakefulness, respectively, preceding task performance. Five out of 20 subjects (25.0%) gained insight in each of these two conditions, a level nearly identical to that of the wake conditions of the main experiment and substantially lower than in the sleep condition of the main experiment (χ2 = 7.07, d.f. = 2, P = 0.029; Fig. 2; the same result, P < 0.01, was obtained in this comparison when, as a control for overall task practice, the last rather than the first 10 of the 13 blocks were analysed). Thus, compared to subjects lacking initial training, subjects of the main experiment benefited from the 8-h post-training interval only if it was spent asleep. Critically, the sleep group of the main experiment also performed substantially better than subjects tested in the early morning after sleep without having established a task representation before sleep, ruling out an unspecific proactive influence of sleep on subsequent task performance.
Insightful behaviour can announce itself by characteristic antecedents19,20. Our task allowed us to identify sleep-related behavioural antecedents of insight in characteristic changes of reaction time patterns across the sleep and wake intervals in the main experiment. Following previous work with the NRT18, we analysed the subjects' seven reaction times of each response string separately for the initial digit (response 1), the following three digits undetermined by the task structure (responses 2–4), and the last three digits (responses 5–7), which were fully determined because they were always mirroring responses 2–4 (Fig. 1a). We identified changes in the three response types from the last block of initial training to the first block of retesting separately for ‘solvers’ (subjects who later gained insight into the hidden rule) and ‘nonsolvers’ (subjects not gaining insight). Overall, reaction times decreased across both sleep and wakefulness. However, this decrease differed saliently between solvers and nonsolvers in the sleep and wake conditions (the data of the two wake conditions did not differ and were combined in this analysis). Notably, in the sleep condition the solvers showed only marginal speeding of reaction times across sleep as compared to the profound decrease in the nonsolvers' reaction times (overall speeding 29.4 ± 27.2 ms in solvers versus 206.4 ± 35.1 ms in nonsolvers; F-test variance ratio (F1,20) = 16.35, P < 0.001, for solver/nonsolver main effect; Fig. 3a). This differential influence of sleep was most obvious for the first response, which in solvers even slowed down across sleep (F2,40 = 5.25, P < 0.01, for solver/nonsolver × response type interaction). There were no similar differences between solvers and nonsolvers across the wake intervals, resulting generally in intermediate gains in reaction times (P > 0.44 and 0.14, for respective main effect and interaction; Fig. 3b). Independent of the sleep/wake conditions, the first response in later solvers as the only response was delayed compared to nonsolvers already in the last block of the initial training (F2,124 = 6.91, P < 0.005, for response type × solver/nonsolver interaction). The slowing of the initial response might be related to processes of search and task analysis21, originating from an incipient representation of the hidden rule. Becoming particularly pronounced at retesting after sleep, this slowing could thus mean that sleep amplifies previously germinated antecedents of insight.
In conclusion, our results show that sleep acts on newly acquired mental representations of a task such that insight into hidden task structures is facilitated. Control conditions excluded the idea that this effect was due to non-specific effects of sleep deprivation, circadian rhythm, or proactive influences of sleep on subsequent capabilities of problem solving or divergent thinking22,23,24. Our reaction time data argue against the view that the qualitative change evolves from a sleep-dependent strengthening of procedural (implicit) memories, which in sequential motor tasks expresses itself in distinctly accelerated reaction times25. Here, sleep profoundly accelerated reaction times only in nonsolvers, but not in solvers, indicating that restructuring originates from an effect of sleep on memory representations different from those underlying procedural motor learning. Specifically, the slowing of reaction times in solvers appears to reflect the presence of an incipient representation of the rule overlapping with that required for implicit task performance. By amplification during sleep, this novel representation could eventually start to dominate the implicit memory representation, a process expressing itself in insightful behaviour as a consequence of an overall restructured representation.
The task representations associated with sleep-dependent gain of insight may be restructured by activity of the hippocampus and related medial temporal lobe structures, which, in connection with prefrontal cortical areas, are considered to play an essential role for generating awareness in memory26,27,28. Reactivation of hippocampal cell assemblies during sleep6,7,8 is regarded as a mechanism by which recently encoded materials stored temporarily in autoassociative hippocampal networks are played back to the neocortex where they are gradually incorporated into preexisting knowledge representations15,29. This process of incorporation underlying long-term storage of previously acquired memories then forms the basis for a remodelling and qualitative restructuring of memory representations. Thus, our data support the concept that sleep, by hippocampal-neocortical replay, not only strengthens memory traces quantitatively, but can also ‘catalyse’ mental restructuring, thereby setting the stage for the emergence of insight.
Sixty-six healthy subjects (age 18–31 yr) recruited at the University of Lübeck participated in the main experiment (twenty-two in each condition) and forty (age 20–32 yr) in the supplementary experiment (twenty in each condition). There were equal numbers of women and men in all five groups. The Number Reduction Task (explained in Fig. 1a) was adopted from ref. 18. Before the experiment proper, subjects had to perform without mistakes on ten practice strings to assure correct understanding of the ‘same’ and ‘different’ rule. Each task block consisted of 30 trials.
The hidden rule was abstract, that is, dependent on relational patterns rather than on fixed stimulus–stimulus or stimulus–response repetitions as in classical conditioning or in typical serial reaction-time tasks. In principle, insight into the hidden rule could be gained in different ways, which, however, were behaviourally equivalent and not treated separately here. The gain of insight reduced the total time to reach the final solution for a string abruptly from 8.73 ± 0.59 s to 2.39 ± 0.17 s. Post-experimental questionnaires confirmed the gain of explicit knowledge of the hidden rule in all subjects identified as solvers on the basis of their behavioural change.
In the design of the main experiment (Fig. 1b), three blocks were chosen as an initial training period based on previous work17,18 and on pilot experiments indicating that at this level of task difficulty only a few subjects were able to recognize the hidden rule within this early phase. Here, five subjects gained insight into the hidden rule already during initial training and, thus had to be replaced by additional subjects. To exclude that insight into the hidden rule had taken place spontaneously between initial training and retesting, subjects were asked in a control questionnaire before retesting whether any thoughts or mentations regarding the task had occurred after initial training. No subject reported any relevant thoughts or dreams pertinent to the task.
Subjects in the sleep condition of the main experiment spent a habituation night in the sleep laboratory before participation. Sleep was monitored by standard polysomnography including electroencephalogram (from left and right central electrodes), vertical and horizontal electrooculogram, and electromyogram (from chin electrodes). Recordings were scored off-line according to standard criteria30, revealing normal sleep architecture for the experimental nights (total sleep time, 480.5 ± 3.9 min; sleep onset, 12.9 ± 1.5 min; wake, 0.2 ± 0.1%; sleep stage 1, 4.5 ± 0.7%; sleep stage 2, 58.6 ± 1.8%; slow-wave sleep, 17.2 ± 1.5%; rapid-eye-movement sleep, 17.7 ± 1.2%).
Reaction time analyses across periods of sleep versus wakefulness (Fig. 3) were performed for correct response strings by a global 2 (sleep/wake) × 2 (solver/nonsolver) × 3 (response type) analysis of variance (ANOVA) and subsequent separate 2 (solver/nonsolver) × 3 (response type) ANOVA for sleep and wake subjects. Pairwise contrasts were specified by t-tests for statistically significant main effects and interactions. Degrees of freedom were adjusted using the Greenhouse–Geisser correction.
We thank S. Sabban, C. Benedict, M. Degirmenci, L. Hecking, A. Otterbein, and M. Rose for their assistance. This work was supported by grants from the Deutsche Forschungsgemeinschaft to J.B. and R.V.