A trainee first solves a novel type of problem by reading each instruction, naming the rule and checking every intermediate step. Weeks later, the same trainee can respond quickly while scarcely describing the route taken. The visible change looks like proceduralization, yet speed alone cannot establish what happened. The trainee might have memorized repeated answers, learned to ignore irrelevant information, or changed strategy. A defensible study must distinguish these explanations.
Measuring proceduralization asks whether performance has become supported by task-specific procedures rather than by repeatedly consulting explicit instructions. It is a question about process inferred from patterns of behavior, not a property directly read from a stopwatch. This methods entry describes the contrasts that can make the inference plausible, including strategy reports, transfer tests, accuracy, dual-task costs and item specificity. It also shows why a learning curve or feeling of ease cannot by itself identify a mechanism.
Define the proposed transition
Anderson’s account of cognitive skill acquisition distinguished an early declarative stage from later procedural performance. In that framework, rules represented as facts can become embedded in productions used to perform the task. The theory provides a precise candidate mechanism. A study that uses the word proceduralization should say whether it means this model-specific compilation process, a broader shift away from explicit step-by-step guidance, or simply improved performance after practice.
These meanings overlap but are not equivalent. A person can remember a practiced answer without compiling a general procedure. They can use a procedure while still being able to explain it. They can become faster because they skip unnecessary checks, not because the original rules were compiled. The observation must therefore be matched to the claim. “Participants solved trained items faster” is a result; “they proceduralized a rule” is an interpretation needing additional contrasts.
The target skill should be specified at a useful grain. In arithmetic, is the procedure applying a formula, retrieving a fact or choosing among problem types? In clinical reasoning, is it identifying a cue, combining cues or deciding when to seek more evidence? A total score over unlike items may improve while the particular component of interest does not. Item-level analysis and carefully chosen novel cases can show what changed.
Training history is part of the definition. A procedure learned under consistent inputs may work only for those inputs. Taatgen’s account of skill and transfer treats the scope of learned productions as an empirical question. If a person performs quickly on familiar examples but fails when surface details change, the observed change may be narrow. That can still be a genuine learning effect; it should not be reported as broad procedural mastery.
Learning curves are a beginning, not a mechanism
Many practice studies plot response time against trial number. A downward curve is easy to see and useful to report, but several mechanisms can produce it. Participants may become familiar with the apparatus, learn the response key, anticipate repeated stimuli, change strategy or abandon unneeded information. Averaging across participants can also smooth abrupt individual strategy shifts into a gradual group curve. A curve alone does not reveal what was learned.
Anderson’s framework discussed practice-related speedup as one pattern compatible with compilation. Logan’s instance theory offered a different account in which repeated exposure creates memory instances that can be retrieved rapidly. Rickard’s work on pseudoarithmetic examined shifts from multistep algorithms to memory-based strategies. These theories differ in what the learner stores and retrieves. A careful methods paper uses that disagreement to design discriminating tests rather than treating faster performance as confirmation of whichever theory it prefers.
Report both speed and accuracy. A trainee might answer faster by guessing more, accepting narrower evidence or omitting a required safety check. A learning curve showing only latency can mistake a change in caution for a gain in skill. Plot error types and response-time distributions, not merely mean correct-trial speed. Excluding all errors before analyzing speed can hide the very tradeoff that changed.
Practice effects can also be item specific. A learner may be fast on exact repeated inputs because answers have become familiar. Include untrained items that require the same rule, and altered items that reveal whether the person still checks relevant conditions. If improvement transfers to structurally matched novel inputs, the case for a reusable procedure is stronger. If it does not, the right conclusion may be efficient recall of practiced cases.
Observe strategy without assuming reports are complete
Ask participants how they solved representative trials, especially before and after practice. A report of reading the rule each time early in training and later using a direct response is relevant evidence. It remains a self-report: people may reconstruct a plausible story after answering, omit fleeting steps or lack vocabulary for a learned routine. A strategy interview should be paired with behavior that the proposed strategies predict.
One option is to use diagnostic items. If an algorithm requires several operations, vary the number of operations while keeping the response format stable. A shrinking relation between step count and response time may indicate a changed route. But it could also arise from shortcut recognition or item memory. The researcher must consider what alternative strategies each probe allows.
Another option is to interrupt or alter information at different points. A learner using explicit instructions may need the rule card visible, while a learned procedure may be less dependent on it. Yet removing a card can introduce anxiety or memory demand, and a highly practiced algorithm can still be articulated. Compare conditions that differ in a clearly specified dependency rather than declaring an unobservable stage from one manipulation.
Think-aloud methods deserve special caution. Speaking during a task can slow it and change attention. A person may revert to explicit steps because the experimenter asked for an explanation. Retrospective probes avoid that disruption but increase reconstruction. Researchers can sample strategy on separate trials, use brief trial-level choices about method, and check whether those choices align with predicted timing and error patterns. No single report is the hidden process itself.
Dual tasks and attention costs
A learned routine often appears less demanding than a novel one. Researchers may add a secondary task to test whether primary performance is less disrupted after practice. This can provide evidence about resource demands, but not a pure test of proceduralization. Participants can reallocate attention, prioritize one task over another or simplify the primary response. Secondary-task performance must be recorded too.
Use a secondary task that does not simply interfere with the primary response modality. Asking a person to tap while learning a motor sequence can create motor competition unrelated to the proposed cognitive process. If a rule task depends on verbal rehearsal, a verbal secondary task may be informative, but the interpretation is about interference under that combination. Different secondary tasks can expose whether the apparent automaticity is broad or resource specific.
Improved dual-task performance after practice could reflect more efficient execution, direct retrieval, reduced checking or familiarity with dividing attention. The contrast is therefore supportive, not decisive. A strong study combines it with novel-item transfer, strategy evidence and error analysis. The aim is to rule out enough alternatives for a bounded conclusion, not to discover a single magic marker of proceduralization.
Information reduction as a rival account
Haider and Frensch argued that practice can change which parts of a task people process. In their experiments, participants learned to focus on task-relevant information and neglect redundant material. That mechanism can make performance faster without showing that all explicit knowledge has been compiled into procedures. It is especially relevant when stimuli contain distracting or repeated features.
A design can test this by changing the status of formerly irrelevant information. If learners ignore a feature that becomes diagnostic, errors may reveal what they had stopped processing. Eye movement or response-pattern measures can help identify attention allocation, but they too require interpretation. A gaze shift is evidence about sampling of information, not a direct record of a rule representation.
Information reduction need not be an opponent of proceduralization in every theory. A compiled procedure may itself specify which cues to inspect. The methodological point is that a latency reduction can be explained by multiple changes. The researcher should predefine which pattern would distinguish the favored account and report when the data support a family of explanations rather than one exclusive mechanism.
Transfer and flexibility test the boundary
A procedure useful beyond repeated trials should survive at least some variation. Transfer tests can change surface form while preserving structure, introduce a new response format or combine familiar components in unfamiliar ways. The distance of transfer matters. A new item with one changed symbol is different from a genuinely new problem context. Report exactly how the test differs from training and what prior knowledge is required.
Taatgen’s work on the nature and transfer of cognitive skills illustrates why generality must be modeled rather than presumed. A production learned for one task can be reused when its conditions match another, but the fit is a theoretical claim to test. A researcher should not call a narrow success “general skill” simply because both tasks share a broad label such as reasoning or mathematics.
Flexibility can also decline. A routine may become fast and accurate on trained conditions yet be slow to adapt when a relevant rule changes. A rule-reversal or exception test can reveal whether the learner notices and updates. Failure under changed conditions is not always evidence of poor learning; it may show the cost of a highly tuned procedure. Report both efficiency in the original environment and adaptation to change.
Retention is another boundary. A gain measured immediately after many repeated trials could fade, while a slower training condition may support later performance. Include delayed tests when the claim concerns durable procedural skill. Even then, delay alone does not distinguish a compiled rule from stable instance memory. Pair retention with structural transfer and process-sensitive probes.
Analysis and reporting choices
Prespecify the unit of practice, the predicted signature and the rival mechanisms. Plot individual trajectories as well as group averages. Report response-time distributions, accuracy and error categories on trained and novel items. State whether trials were identical, varied within a rule family or drawn from multiple families. These details determine what a speed gain can mean.
Model baseline differences and prior experience. Participants who start with more familiarity may already use an efficient method, leaving less room for change. Others may improve quickly because they discover a shortcut. A group average can hide these different routes. Strategy classification should be reproducible, with coding rules and agreement reported when human observers are involved.
Where possible, compare two theories with predictions that genuinely differ. For example, a production-compilation account and item-specific retrieval account may make different predictions for structurally new items. A comparison has value only if the novel items are matched in difficulty and do not introduce unrelated demands. Failure to transfer after a poorly designed test is not decisive evidence for one account.
Finally, avoid naming the mechanism at a level more precise than the evidence. It may be appropriate to conclude that training reduced reliance on visible instructions and improved transfer. It may be premature to conclude that a particular production rule was compiled in the brain. A good report makes the ladder from observation to interpretation visible.
What the evidence does not show
The cited work does not provide a universal biomarker of proceduralization. Faster responses, smoother performance, reduced verbal explanation, smaller dual-task costs and changed neural activity can each arise through more than one route. Anderson’s model is influential, but Logan’s instance theory and information-reduction findings show why the same broad practice curve can admit different mechanisms. No one task establishes that declarative knowledge has disappeared.
Nor does proceduralization imply that a person cannot explain a skill or that explicit instruction becomes useless. A trained performer may use a fast routine most of the time and switch to deliberate reasoning when an exception appears. The evidence does not justify ranking people by a single speed score or assuming that a practiced routine transfers to every related domain. General evidence of skill-related plasticity is not a direct measure of which procedure a person used on a task.
The measurement framework offered here is Institute synthesis from the cited models and experiments, not a newly validated protocol. Its purpose is to make claims testable and proportionate: define what changed, what remained stable, and which alternative process could have produced the same observation.
Return to the trainee
The trainee’s rapid response might be an efficient procedure. To find out, ask whether the response remains accurate on new problems that require the same structure, whether performance survives removal of the instruction sheet, which errors appear when conditions change, and whether attention is still allocated to the necessary cues. Record how these results evolve across training rather than deciding from the final stopwatch reading.
Proceduralization is most informative as an inference about how a task is carried out. The strongest evidence comes from a pattern of contrasts: speed with accuracy, trained with novel cases, explicit guidance with independent performance, and stable conditions with exceptions. That pattern can support a careful conclusion while leaving room for multiple learning mechanisms when the data do not separate them.
Research references
- Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89(4), 369–406. View source.
- Logan, G. D. (1988). Toward an instance theory of automatization. Psychological Review, 95(4), 492–527. View source.
- Rickard, T. C. (1997). Bending the power law: A CMPL theory of strategy shifts and the automatization of cognitive skills. Journal of Experimental Psychology: General, 126(3), 288–311. View source.
- Haider, H., & Frensch, P. A. (1996). The role of information reduction in skill acquisition. Cognitive Psychology, 30(3), 304–337. View source.
- Taatgen, N. A. (2013). The nature and transfer of cognitive skills. Psychological Review, 120(3), 439–471. View source.