Does Cognitive Training Transfer? The Evidence for Process Training

Does getting better at one mental task make you better at others? The whole premise of cognitive training rests on that question, and the honest answer arrives in two halves that sound like a contradiction. Most "brain training" won't make you broadly smarter — the skeptics have earned that conclusion. But the same studies that puncture the inflated claims contain a quieter, sturdier result, and that result doubles as a blueprint for the kind of training that does transfer.

Start With the Bad News

Commercial "brain training" deserves much of its bad reputation. The strongest claims — that a few weeks of puzzles will make you globally smarter — have repeatedly failed to replicate. Two careful reviews make the point cleanly. Melby-Lervåg, Redick, and Hulme (2016) re-analyzed the working-memory training literature and found that, once you restrict attention to studies with active control groups and proper baselines, the evidence for far transfer — gains on intelligence, reading, or arithmetic from training an unrelated task — shrinks toward zero. Simons and colleagues (2016), in an exhaustive consensus review, reached a similar verdict: most brain-training products have little good evidence for broad real-world benefit.

These reviews get summarized as "cognitive training doesn't work." That isn't what they found. They found something narrower: training a task improves that task and its close relatives, and stops there.

What Near Transfer Buys You

The same literature that fails to find far transfer reliably finds near transfer. Train working memory and you improve on other working-memory tasks — including ones you never practiced, as long as they draw on the same underlying capacity. Meta-analyses put this near-transfer effect in the moderate-to-large range (commonly around half a standard deviation, sometimes more, immediately post-training). The effect is real, and it is bounded: it stops at the edge of the trained process.

Process Overlap Theory predicts exactly that shape. Its claim is that mental tests correlate not because of a single "general intelligence" sitting behind them, but because they draw on overlapping processes — so strengthening one process lifts every task that samples it, and only those. (I make the full case for it in Process Overlap Theory: Rethinking Intelligence and Cognitive Training.) Far transfer was always the wrong thing to ask for; it expected training to leap across processes it never touched. Near transfer to a foundational process asks for something simpler: improve the process, then collect the benefit wherever the process is used. "Near transfer yes, far transfer no" and "train the shared processes" are the same claim approached from two directions.

The question that matters, then, isn't whether training produces far transfer. It's how foundational the trained process is, and how many tasks draw on it. Train something narrow and the benefit stays narrow. Train something that sits underneath reasoning itself, and that same bounded near transfer reaches a long way.

That only holds if some process is both trainable and foundational enough to move general reasoning. Is there one?

The Existence Proofs: When IQ Does Move

There are trials — smaller and less famous than the headline brain-training studies — where fluid intelligence genuinely moved. Two stand out: they came from overlapping teams, and they share a feature the failed studies lack.

Bor et al. (2014). A nine-week regime trained 14 adults to associate 13 letters with specific colors. The remarkable part isn't the letters — it's that the training was, start to finish, intensive color memory: holding a precise hue in mind across a delay and reproducing it, with the foils stepped ever closer in color space as participants improved. The headline result for our purposes was peripheral to the authors' own goal: the trained group gained about 12 IQ points on a culture-fair fluid-intelligence test (≈0.8 SD; Cohen's d = 1.21), while a control group taking the same test twice gained nothing. I unpack this study in depth in Color and Cognition; here it's one of two data points.

Rudebeck, Bor, et al. (2012). A spatial dual n-back: participants monitored real-world scenes appearing across eight frames in a 3D room, tracking both which scene and which location repeated. After 20 days, trainers improved significantly on a matrix-reasoning test of fluid intelligence (BOMAT) relative to non-trainers (d ≈ 0.7–1.0). Crucially, the 440 scenes were "specifically picked so that they could not be easily encoded verbally."

Two different labs, two different surface tasks — color recall and spatial n-back — both moving fluid intelligence. The question is what they have in common that the failed studies don't.

What Those Studies Share: Non-Verbalizable Practice

Both trained the holding of visual information in its pure form, one that can't be put into words.

In the Bor regime, the workhorse tasks were delayed match-to-sample: see a color, hold it across a few seconds of blank screen, then pick it out. As accuracy rose, the distractor colors crept 20% closer in RGB space each time — driving discriminations far finer than any verbal label ("blue," "green") can carry. You cannot solve "was it this blue or that blue, three seconds ago" by storing the word blue. You have to hold the percept itself. That delayed-match task showed the largest training gain of anything in the study.

In the Rudebeck regime, the scenes were chosen, deliberately, to defeat verbal encoding — you can't compactly say "the indoor hallway from two frames ago" the way you can rehearse a digit string. The only way to track repeats is to hold the visual scene.

In the color article I call this imagery training, using "imagery" loosely to mean non-verbalizable internal visual representation. The claim there is that strengthening this representational capacity — the ability to generate and hold a vivid mental percept — is what drove the IQ gains, because mental simulation of exactly this kind is used throughout fluid reasoning: spatial manipulation, mental rotation, holding abstract structure in mind.

This is also where the usual deflationary reading runs out. Bor's authors noted that "the working memory aspects, and not the synesthetic features" might be what raised IQ — and you could say the same of the spatial study: it was just working memory. But the working memory these regimes trained was non-verbalizable visual working memory. When the thing held in mind is a percept with no verbal handle, "it was working memory" and "it was imagery" name the same training, not two rival explanations.

The Pattern Isn't Just Visual

If the principle is real, it shouldn't be confined to color and scenes. Other lines of work point the same way — training a foundational process hard, in a form that blocks shortcuts, and seeing the benefit appear in a different task that shares the process:

• Relational reasoning → fluid intelligence. SMART (Strengthening Mental Abilities with Relational Training) trains derived relational responding using arbitrary, nonsense terms, and several trials report gains on standardized IQ measures. This is the cleanest test of the principle, because relational reasoning is about as foundational as a process gets.
• Reading acceleration → reading fluency. Reading-acceleration training (pushing reading speed past the comfortable rate) has produced fluency gains reported to replicate across independent international samples — a near-transfer effect at the level of the reading-rate process rather than any specific text.
• Long-term-memory retrieval → reading fluency. Training retrieval from long-term memory has been reported to transfer to reading fluency — again, transfer along a shared process rather than a shared task.

None of these is individually decisive. Taken together with the imagery trials, they sketch a consistent shape: transfer happens, but it travels along processes, surfacing in tasks that look unrelated but share the machinery you actually trained.

The Common Thread: No Shortcut Allowed

The successful cases share one feature, and it's exactly what the failures lack: there is no way to cheat them.

The reason vague "brain games" fail is that the brain is lazy in a specific way: given any task, it will find the cheapest strategy that raises the score, and a cheap strategy is almost always task-specific. You get better at the game, not at the process the game was supposed to tax. Abstract training — nonsense terms, content-free relations — exists precisely to remove the cheap strategy, leaving the process as the only thing that can improve.

Non-verbalizability is the same move, in the visual domain. Verbal recoding is the cheap shortcut for visual memory: instead of holding the percept, you store a word and reconstruct. Force the discrimination finer than any word can capture, and you close that exit. What's left to improve is the perceptual representation itself. Imagery training is a special case of abstract, then: block the verbal shortcut so the visual process has to carry the load.

That's the unifying principle behind every case above. Pick a foundational process. Train it across enough variation that no content sticks. Strip the task of any handle that lets the brain cheat. Then the only path to a higher score runs through the process you wanted all along — and the near transfer follows it wherever it goes.

What Would Settle It

The caveats are real:

• Small samples. The imagery trials had 14 and ~27 trainers. These are existence proofs, not population estimates.
• Passive controls. Both compared trainers to people who did nothing, so a motivation or expectancy difference can't be fully ruled out — the very critique the 2016 reviews leveled at the field. The authors of both studies say as much and call for active-control replications.
• A dose-response wrinkle. In the Rudebeck study, it was pre-training ability, not how much someone improved on the n-back, that predicted the fluid-intelligence gain. A clean "more training → more IQ" story would have predicted the opposite, so the mechanism isn't fully pinned down.
• The broad averages stay low. With active controls, the average far-transfer effect across the WM-training literature is small to nil (Au et al. 2015 put n-back → Gf at g ≈ 0.24; Melby-Lervåg et al. 2016 put it lower). The strong cases are specific regimes, not the genre as a whole.

What would settle it is obvious and overdue: active-control trials of non-verbalizable, mastery-based process training, powered to detect a fluid-intelligence effect, run by teams without a product to sell. The existing data is enough to make the bet worth taking. It is not enough to call it closed.

Why We Train the Way We Do

This is the evidentiary spine of Relatoria. We don't build games and hope for far transfer; we try to identify foundational processes and train each one in the most shortcut-proof way we can. Frames trains relational reasoning on the SMART model — abstract, nonsense relations with nothing to memorize. MemoHue trains exactly the non-verbalizable color memory the Bor regime used, with the distractors closing in across CIE LUV color space so the discriminations stay finer than words. The bet isn't that any one task makes you smarter. It's that strengthening a genuinely foundational process pays out as near transfer everywhere that process is used — which, if Process Overlap Theory is right, is nearly everywhere.

References

• Process Overlap Theory: Kovacs, K. & Conway, A.R.A. Process Overlap Theory: A Unified Account of the General Factor of Intelligence. Psychological Inquiry 27, 151–177 (2016). DOI: 10.1080/1047840X.2016.1153946
• Far transfer is weak (meta-analysis): Melby-Lervåg, M., Redick, T.S. & Hulme, C. Working Memory Training Does Not Improve Performance on Measures of Intelligence or Other Measures of "Far Transfer." Perspectives on Psychological Science 11, 512–534 (2016). DOI: 10.1177/1745691616635612
• Brain-training consensus review: Simons, D.J. et al. Do "Brain-Training" Programs Work? Psychological Science in the Public Interest 17, 103–186 (2016). DOI: 10.1177/1529100616661983
• Near transfer (meta-analysis): Melby-Lervåg, M. & Hulme, C. Is working memory training effective? A meta-analytic review. Developmental Psychology 49, 270–291 (2013). DOI: 10.1037/a0028228
• n-back → Gf (meta-analysis): Au, J. et al. Improving fluid intelligence with training on working memory: a meta-analysis. Psychonomic Bulletin & Review 22, 366–377 (2015). DOI: 10.3758/s13423-014-0699-x
• Imagery trial (color): Bor, D., Rothen, N., Schwartzman, D.J., Clayton, S. & Seth, A.K. Adults Can Be Trained to Acquire Synesthetic Experiences. Scientific Reports 4, 7089 (2014). DOI: 10.1038/srep07089
• Imagery trial (spatial): Rudebeck, S.R., Bor, D., Ormond, A., O'Reilly, J.X. & Lee, A.C.H. A Potential Spatial Working Memory Training Task to Improve Both Episodic Memory and Fluid Intelligence. PLoS ONE 7, e50431 (2012). DOI: 10.1371/journal.pone.0050431