Predictive Processing on Covert Vs. Overt Retrieval
A response to "Does Thinking Silently help Students learn" a blog post by Carl Hendrick
Retrieval Practice Through the Lens of Predictive Processing: Why Overt Beats Covert
Carl Hendrick recently summarised a compelling meta-analysis (Yu et al., 2025), in a Substack post titled Does Thinking Silently Help Students Learn?. He highlights that while covert (silent) retrieval does improve learning, it is only effective when followed by feedback. Without feedback, the benefit almost disappears. Hendrick also explains that overt retrieval - speaking, writing, or typing - consistently outperforms covert strategies. He posits that this is likely because it creates a 'truncated search', where students settle for half-formed answers in their heads. He argues that simple practices like cold calling can turn mental effort into durable learning if paired with feedback and time for consolidation.
I will review the covert vs. overt findings through the lens of Predictive Processing and Active Inference theory. From the perspective of these latest cognitive science models, these findings are not just unsurprising, they’re expected.
What is Predictive Processing / Active Inference?
For readers unfamiliar with Predictive Processing and Active Inference, these are the latest unifying frameworks emerging from cognitive and neuroscience. They are rapidly becoming the dominant models for explaining cognition and consciousness, extending into fields such as psychology, psychiatry, artificial intelligence, and functional disorders like persistent pain (notably, education and the science of learning remain underrepresented in this conversation!)
These theories view the brain not as a passive processor of incoming information but as a dynamic prediction machine, constantly generating hypotheses about the world and updating them based on mismatches between expectation and reality, known as prediction errors. For an education-focused introduction, see my Substack post: “Teachers Are Prediction Error Managers”. For wider context, explore the following introductory talks: Is Reality a Controlled Hallucination? - Anil Seth, How the brain shapes reality - Andy Clark, and Active Inference in the Brain -Karl Friston.
Covert Retrieval and Prediction Error
When students engage in covert retrieval, they generate internal predictions about stored information. But without feedback, without a persistent, unresolved prediction error, their brain receives no signal that the retrieved information might be inaccurate. The generative model remains unchallenged and potentially reinforces incorrectly embedded parts of the generative model. The brain, in essence, thinks: “I retrieved something; prediction succeeded,” regardless of whether it matched the target knowledge.
Predictive Processing tells us that learning requires persistent prediction error, not just error, but persistent and salient error that can't be explained away. Feedback is the critical ingredient that flags the mismatch, forcing an update of the generative model. Without it, covert retrieval risks becoming a rehearsal of confident but flawed predictions.
Furthermore, in the absence of feedback, the brain is likely to assign low or medium precision to the recalled information, especially if the student is unsure. This means the retrieval has little impact on strengthening or updating the generative model. But it can be worse: if a student is confident but wrong, the brain assigns high precision to an inaccurate prediction, reinforcing a faulty model. This is how misconceptions become embedded: when incorrect predictions are treated as successful in the absence of an error signal.
Overt Retrieval and Active Inference
Cold calling deserves particular attention here. From a Predictive Processing perspective, it works not just because it increases attention, but because it increases precision weighting—students assign higher importance to the retrieval prediction when they know it may be immediately tested. This anticipation of action heightens the salience of the prediction error that follows. In other words, when a student knows they might be called upon, they are more likely to generate a full retrieval attempt and hold it up against the eventual feedback. This creates the conditions for effective generative model updating.
Cold calling also promotes what Hendrick describes as a shift from passive observation to active mental construction. When everyone prepares to respond, each student engages their own prediction machinery, even if they aren't ultimately called on. If feedback is then provided, either through a peer's answer or a revealed correct response, every student has the opportunity to compare and calibrate their internal prediction, embedding the learning more robustly.
However, for all the students who are not the person called on, cold calling is still a covert retrieval.
In contrast, overt retrieval doesn’t just test stored content, it activates the entire perceptual-action loop. A student must coordinate a cascade of predictions:
From sensory cue (e.g., a question or image)
To semantic interpretation
To motor planning (writing/speaking)
And finally to external action
This full generative loop, from cue to action, mirrors the demands of real-world retrieval. It strengthens the generative model not just in terms of what to retrieve, but how and when to express it. In Predictive Processing terms, it's not just learning the content, but updating the precision-weighted mappings between perceptual input and active output, crucial for generalisation and flexible use.
This is why techniques such as using mini whiteboards, "all write" strategies, or "turn and talk" can be more effective than cold calling for engaging the entire class in the perceptual-action loop. These approaches require every student to externally produce a response, forcing the full generative process of retrieval, formulation, and action. This time from all learners, not just the one who is called upon. In Predictive Processing terms, they offer broader opportunities for updating each student's generative model, as all students commit to a prediction and receive corrective feedback in real time. This makes such strategies especially valuable for classroom learning that aims for high-fidelity generative model refinement across the group.
Feedback and the “Beyond First Success” Effect
It struck me that these covert vs. overt differences in learning align with what Matt and Femi described as the "First Success" effect in a recent episode of their podcast First Success. In it, Matt reflects on his experience of finally succeeding at the complex manoeuvre of gybing on a wing foil board, despite having watched videos, listened to explanations, and observed others. It was the first successful execution that shifted his learning dramatically.
This applies directly to education. Whether it's a teacher successfully pulling off a well-structured “turn and talk,” or a student solving their first quadratic, the same dynamic is at play. In terms of Predictive Processing and Active Inference, that first success is not just symbolic--it marks the first time the learner’s prediction machine has experienced a full sensory, affective, and action-based prediction sequence that ends successfully. It completes and grounds what had previously only been observed or imagined.
The generative model—up to that point scaffolded by instruction, modelling, or observation—is now pencilled in with a rich internal map of what it feels like to succeed. Subsequent attempts are no longer hypothetical; they can be compared against this real, embodied reference. The learner can now better isolate which elements of the sequence need adjustment, giving greater precision to each new iteration. With repetition, this pencilled-in experience becomes inked in—a robust, generalisable model built from successful prediction-error resolution.
Matt and Femi also highlighted that part of the power of the wing foil gybing first success was that the success was unmistakable—it was clear whether the manoeuvre had been executed correctly or not. In classroom practice, however, the success of a teaching strategy, routine, or technique is often not as immediately observable. This may help explain why instructional coaching models such as responsive coaching can be so effective. These approaches provide the feedback loop needed to confirm success and drive updates in the teacher’s generative model. It also sheds light on the value of defining 'Success Criteria' in cognitive coaching, as discussed by Adam Kohlbeck and Sarah Cottingham.
So, a "first success", with a clear feedback of success, allows a recalibration of the entire prediction-action loop, grounded in full sensory experience and capable of guiding future action with increasing fidelity.
So what? - Insight from Predictive Processing / Active Inference.
The key insight from viewing learning through the lens of Predictive Processing and Active Inference is that learning is fundamentally a process of re-engineering our prediction machinery. This machinery is not limited to abstract thought, it is deeply integrated with our sensory experience, emotions, and actions. As a result, the most powerful learning happens when the whole system is engaged.
Learning, understood as the updating of our generative model, only occurs when prediction errors are persistent and meaningful. This means feedback matters enormously. Without some signal of whether an action or response was successful, the brain has no reason to revise its internal models. That’s why it's critical to design learning experiences that provide immediate and clear feedback about success or failure. It is this feedback, this evidence that a prediction worked or didn’t, that drives meaningful change in the learner’s mind and behaviour.
In Summary:
Overt retrieval isn't just more effective by chance; it reflects deeper engagement with the full perception-action loop that Predictive Processing/ Active Inference sees as central to learning. Feedback is vital because it introduces the prediction error signal needed for real updating. Covert retrieval can help, but only if it's anchored in error-sensitive feedback that the brain can use to revise its models.
Acknowledgement
I am dyslexic and I utilise ChatGPT to help me structure my posts, and to have readable grammar and spelling. The content is my own thoughts, unless referenced to others. I review and edit the content before publication.
References
Clark, A. (2024). The experience machine: How our minds predict and shape reality. Random House.
Cognitive Coaching. (2024). Coaching cut 3: Wrap-around success. Retrieved from https://open.substack.com/pub/cognitivecoaching/p/coaching-cut-3-wrap-around-success
Friston, K., Parr, T., & Mirza, M. (2020). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
First success [Podcast episode]. In Beyond Good: A podcast for teachers. Buzzsprout. Retrieved from https://www.buzzsprout.com/1959070/episodes/17208167-first-success
Hendrick, C. (2025). Does thinking silently help students learn? The Learning Dispatch. Retrieved from https://carlhendrick.substack.com/p/does-thinking-silently-help-students
Yu, Y., Zhao, W., Li, A. et al. (2025). Is Covert Retrieval an Effective Learning Strategy? Is It as Effective as Overt Retrieval? Answers from a Meta-Analytic Review. Educational Psychology Review, 37, 52. https://doi.org/10.1007/s10648-025-10024-4
Seth, A. (2021). Being You: A New Science of Consciousness. Faber & Faber.
Wray, A. (2025). Teachers are prediction error managers. Predictably Correct. Retrieved from https://predictablycorrect.substack.com/p/teachers-are-prediction-error-managers
Wray, A. (2025). What are we really talking about when we say “attention”? Predictably Correct. Retrieved from https://predictablycorrect.substack.com/p/what-are-we-really-talking-about