Power of instructions for task implementation: superiority of explicitly instructed over inferred rules

In the following paper, we adopt the term ‘RITL’, but do so without commitment to its specific instantiation in Cole’s papers, but instead, more broadly, as a class of paradigms, all which examine instructions-based performance.

Originally, the NEXT phase was introduced to measure automatic effects of instructions (Meiran et al., 2017), which was not the focus of the current work (though it is involved in Experiments 1 and 2, but not in Experiment 3). This additional process (responding to a NEXT target) might raise the question of whether RITL in its original meaning is measured in this task. We argue that it is. Specifically, perhaps early works concerning RITL focused merely on direct and immediate instructions (e.g., Cole, Bagic, Kass, & Schneider, 2010), but recent developments in this literature (Cole et al., 2017) consider additional forms of RITL tasks, including those involving delayed implementation. In addition, instructions in real life are often performed with some delay and not immediately (reconsider the driving directions example). In our view, the broad definition of RITL (“the ability to rapidly perform novel instructed procedures”, Cole et al., 2017, p.2) qualifies the GO phase of the NEXT paradigm as measuring RITL (see also Cole et al., 2013, Fig. 1b for a non-verbal RITL example that closely resembles our NEXT instructions). More specifically, Cole et al. (2013) defined different forms of RITL, of which the NEXT paradigm should be considered as measuring concrete and simple non-verbal RITL; while a later elaboration of this definition suggests that S–R mappings in the NEXT paradigm could be completely proactively reconfigured (Cole et al., 2017), relative to Cole et al.’s (2013) RITL paradigm. Another consideration for including a NEXT phase relates to yet unpublished experiments showing that performance in the GO phase was approaching ceiling when the NEXT phase was omitted (i.e., participants reached an excellent level of performance shortly after the onset of the experiment), suggesting that delaying implementation might prove as crucial to study RITL in this task if one wishes to avoid near-ceiling levels of performance. Nonetheless, in this study we directly measure the importance of the delaying NEXT phase in influencing the efficiency of explicitly instructed/inferred rules (comparing Experiment 2 with Experiment 3).

Another important comment is that, in contrast to null-hypothesis testing, Bayesian inference generally avoids alpha inflation toward accepting H1 in exploratory analyses, since there is no bias in favor of H1, as both H1 and H0 can be accepted (Rouder et al., 2009).

NEXT errors refer to trials in which participants erroneously executed the GO instructions during the NEXT phase (e.g., pressing “left” instead of the spacebar). We did not remove mini-blocks involving a NEXT error since it does not reflect a problem with rule encoding, but rather the opposite (i.e., a reflexive activation of the instructions in an inappropriate context).

We thank anonymous Reviewer 1 for suggesting this analysis.

It should be noted that the RT interaction between Inference Difficulty and Rule-Type was robust when using a different RT cutoff of 3.5 sd per condition and participant, suggesting that perhaps the effect is increased in the high percentiles of the distribution.

We thank Bernard Hommel for raising this important issue and inspiring Experiments 2 and 3.

Direct and indirect instructions refer to instructed and inferred rules, respectively.