Memory tasks without an intelligence league table

Delay

Animal memory studies mostly use four task types: delay, recognition, working demand and generalization. Each measures a different property, so the scores do not add up to an intelligence ranking. A delay task measures how long an animal can hold information across a gap before acting on it. The basic setup shows a subject some cue or sample, removes it, inserts a waiting period, and then checks whether the animal’s later choice matches what it saw earlier.

What counts as “success” is a specific behavioral match at test. The task itself makes no claim about conscious remembering or emotional experience. Where a design varies the delay, check whether the study uses it to examine changes in accuracy. In that design, delay becomes an experimental variable, examined on purpose. It isn’t there as a fixed background condition.

Any point where accuracy drops off is a measured threshold specific to the task, the species, and the testing apparatus used. It doesn’t transfer automatically to a different memory system or a different set of stimuli.

Delay-based designs depend heavily on procedural detail: how the sample is presented, what counts as interference during the gap, how many trials are run. Reporting quality decides whether a result can be interpreted at all. That’s where the ARRIVE guidelines 2.0 come in, organised into an Essential 10 and a longer Recommended set. ARRIVE is a reporting checklist; it does not say what any memory task means.

ARRIVE addresses whether the write-up gives readers enough method detail (sample size, randomization, blinding) to judge if the delay effect is reliable in the first place. Grading whether a memory finding is impressive or biologically important falls outside that job.

That distinction is worth holding onto when comparing tasks across studies. Two papers can both report “successful delayed matching” while differing completely in delay duration, trial count, and control conditions.

A longer delay tolerated in one study reflects a different task than a shorter delay tested elsewhere. Reading it as evidence of a better memory system credits the number with more than the comparison supports. Treating delay length as a single scoreboard number flattens details that the original researchers built in on purpose.

Recognition

A recognition task asks whether an animal treats a previously encountered item or location as familiar, typically by measuring preferential attention or approach toward novelty. The common logic: if an animal spends more time investigating a new object over a familiar one, that pattern is read as evidence the familiar item left some trace.

The measure itself is an exploration ratio or a discrimination score. Reading it as a direct readout of memory strength credits it with more precision than it has.

Delay tasks probe how long information persists. Recognition tasks probe whether a stored representation can be distinguished from something novel at the moment of testing, regardless of how long ago the original exposure happened.

A study can combine both, inserting a longer gap before the novelty test, but the novelty-preference measure and the delay-tolerance measure remain conceptually separate outcomes. Interpretation gets complicated by confounds that have nothing to do with memory, such as how objects were positioned, cleaned and rotated between trials.

Which is why the procedural description carries the weight here. How objects were positioned, cleaned, and rotated between trials counts as real information: it’s what a reader needs to tell a memory effect from a difference in the objects themselves. A write-up that leaves it out stops short of the point where a reader could actually check.

Reporting standards addressed by the ARRIVE guidelines speak directly to this: they ask authors to specify exclusion criteria, randomization, and experimental procedures in enough detail that readers can judge whether a novelty effect is a memory finding or an artifact of setup. Comparing recognition scores across species or studies without checking those procedural details risks comparing different confound profiles and mistaking that difference for one in memory abilities.

A high discrimination score in one lab’s object-recognition setup and a lower one in another’s says something about each specific protocol first, and only tentatively about the animals’ underlying capacities.

Working demand

Working demand refers to how much information an animal must hold and manipulate simultaneously while performing a task. How “smart” the performance looks doesn’t factor into that definition. Tasks that require tracking several items at once, updating a running total, or ignoring irrelevant information mid-trial impose a heavier working-memory load than tasks with a single sample and single delay.

This dimension differs from a pure delay design, where the challenge is duration, and from recognition, where the challenge is discrimination at test. Working demand is about concurrent load: how many pieces of information are active at once and how much interference the animal must filter out.

Demand gets manipulated by adding distractor items, increasing the number of to-be-remembered locations, or requiring updating, which means dropping an old item and replacing it with a new one. Where a given study reports performance falling off under added load, that point of decline is a finding about that species, that apparatus and that trial structure.

It doesn’t establish a general capacity ceiling that applies beyond the tested conditions. Reporting on working-demand studies benefits from the same procedural transparency relevant to delay and recognition work.

Sample size, randomization of trial order, and clear outcome measures (accuracy, latency, error type) all affect whether a claimed demand effect holds up. The ARRIVE guidelines outline exactly this kind of reporting scope: study design and outcome measures. Certifying that a particular task is a valid stand-in for general intelligence sits outside that scope.

Treating a working-demand result as evidence of broad cognitive sophistication overstates what the task measures. The result describes behavior under a specific, engineered load condition, and comparisons across tasks should stay anchored to that specificity.

Generalization

Generalization measures whether a learned response transfers to new but related stimuli. Whether an animal understands an abstract rule the way a person might describe one is a separate question that this measure doesn’t touch. A typical design trains an animal to respond to one stimulus and then tests responses to stimuli that vary along some dimension (shape, color, spatial arrangement).

The transfer pattern observed is the data; any claim about underlying “rule learning” is an interpretation layered on top of that pattern. Generalization sits apart from delay, recognition, and working-demand measures because it targets flexibility of a learned association. Persistence, discriminability, and concurrent load are the separate properties those other three tasks track.

An animal could show strong delayed recall of a specific stimulus and still fail to generalize that response to a slightly altered version of the same stimulus. These are separate outcomes, and a study reporting one doesn’t automatically speak to the others.

Because generalization tests hinge on exactly how “similar” a new stimulus is to the trained one, procedural description of stimulus construction and selection criteria carries particular weight. A gradient of test stimuli described in coarse terms leaves readers unable to judge how demanding the transfer test really was.

The ARRIVE guidelines reporting scope touches this directly, in its emphasis on specifying experimental procedures and inclusion or exclusion criteria clearly enough for outside readers to evaluate the design. Generalization results should stay tied to the tested stimulus set.

A finding that an animal transfers a response across two shapes says nothing certain about transfer across an unrelated sensory dimension, like scent or sound, unless that transfer was tested directly. Broader claims about “flexible cognition” go beyond what a single generalization task can establish on its own.

Can memory tasks rank animal intelligence?

Comparing memory tasks side by side works only when the comparison stays at the level of what each task measures. A ranked score of intelligence is a different kind of claim, one none of these tasks were built to support. Delay, recognition, working demand, and generalization each isolate a different behavioral question, and a single animal or species can perform differently across them without contradiction.

Task type	What it isolates	Typical outcome measure
Delay	Persistence of information over time	Accuracy at increasing delay lengths
Recognition	Discrimination of familiar versus novel	Novelty-preference or exploration ratio
Working demand	Concurrent load and interference	Accuracy or error rate under added load
Generalization	Transfer to related stimuli	Response rate across a stimulus gradient

None of these outcome measures are interchangeable, and none function as a composite intelligence score. A species that tolerates long delays in one published design isn’t shown to be “smarter” than one tested only under high working demand; the tasks aren’t measuring the same thing, so the results don’t stack into a ranking.

Cross-study comparison also depends on whether the underlying reports give enough procedural detail to judge reliability. This is precisely the scope the ARRIVE guidelines address: study design, sample size, randomization, blinding, and outcome measures, laid out so a reader can assess whether a given result is likely to replicate.

That reporting checklist is a transparency tool. It doesn’t endorse any task’s importance, and it doesn’t rank species against each other.

Reading across a set of memory studies with these distinctions in mind means treating each result as an answer to a narrow, specific question. A given task tells you about persistence, discrimination, load tolerance, or transfer, one at a time.

The league table is the thing to refuse. Four tasks measuring four different properties produce four numbers that don’t add, and a ranking across those tasks needs a justification that the individual scores don’t supply.