HSW-0242
A sentence begins in a predictable way. Your mind can lean on context. Then one word arrives that the context did not prepare you for. Processing seems to tighten around it. The unexpected word may be harder to integrate, yet it may also receive unusually careful encoding.
That combination—greater immediate difficulty and potentially more robust representation—looks contradictory only if we assume working memory spends its limited resources evenly.
In a 2026 Journal of Memory and Language paper, Weijie Xu and Richard Futrell proposed Strategic Resource Allocation: a resource-rational account in which working memory preferentially allocates encoding precision to linguistic material that is surprising in context. Their theoretical claim is that, when memory capacity is limited and representations are noisy, unexpected material deserves more precise encoding because it is less recoverable from prior expectations if it is lost. The authors tested predictions of the account using large naturalistic language corpora and reading-time data.
The educational temptation is obvious: make study material surprising and memory will improve. That conclusion is not established by the paper.
The better studying question is narrower and more useful: if memory resources are allocated selectively rather than uniformly, what does that imply about attention, prediction and the way learners should test what survived encoding?
The direct answer
Working memory is constrained. A system facing limited capacity cannot preserve every incoming detail with equal precision. Xu and Futrell’s Strategic Resource Allocation model proposes that one efficient solution is to encode less predictable linguistic input more precisely, because predictable input can be reconstructed more easily from context if its representation becomes noisy.
This is a theory about efficient memory encoding in language processing, supported by corpus and reading-time patterns. It is not evidence that students should add novelty, jokes, colour or surprise to every lesson. Nor does it show that unexpected classroom content is always remembered better.
For studying, the safest application is diagnostic: learners should distinguish what feels predictable because it is genuinely well represented from what feels predictable because context is doing reconstruction work for them. Remove the context later and test the representation itself.
Why equal allocation would be wasteful
Imagine reading: “She spread the warm bread with…” The next word is highly constrained by context. If the final word is butter, partial memory loss may be repairable from the sentence. If the sentence ends with an unusual but meaningful word, reconstruction from prior expectation is harder.
A limited system therefore faces an allocation problem. It can distribute encoding precision uniformly, or it can protect information that will be difficult to reconstruct later. In Xu and Futrell’s resource-rational formulation, prioritising high-surprisal input reduces expected retrieval error under memory constraints.
The key word is expected. The system does not need to preserve every item perfectly. It needs to use finite resources in a way that minimises error overall.
This creates a useful inversion. Material that is harder to predict may require more processing now precisely because it cannot safely be reconstructed later.
Surprisal is a technical idea, not simply “wow”
In language modelling and psycholinguistics, surprisal refers to how unexpected an input is given the preceding context. It is not identical to emotional surprise. A statistically unexpected word can have high surprisal without causing astonishment, and a dramatic fact can feel surprising for reasons that are not captured by word probability.
This matters because educational writing often turns a technical prediction effect into a generic claim that “surprise boosts memory.” HSW-0242 does not make that leap.
The 2026 paper asks how a memory-limited language processor may allocate encoding precision. Its empirical tests concern linguistic dependencies in production and comprehension. Classroom learning adds goals, prior knowledge, conceptual structure, emotion, retrieval practice, feedback and long retention intervals. Those are different layers of evidence.
What the researchers examined
Xu and Futrell studied a phenomenon known as dependency locality: relationships between parts of a sentence tend to become more difficult as intervening material increases. Their Strategic Resource Allocation account predicts that a surprising antecedent should receive a more precise memory representation and therefore be more resistant to interference across a longer dependency.
They examined naturalistic production corpora across ten languages and comprehension data from large reading-time datasets. Across analyses, less predictable antecedents were associated with a reduced locality effect in patterns consistent with the account.
But the findings were not uniform. The production effect was especially evident in Indo-European and head-initial languages, and the authors report important differences across dependency types and languages. That cross-linguistic variability is a major limit on any claim that the proposed principle has already been demonstrated as a simple universal law.
The evidence therefore supports a theoretically interesting efficiency account while leaving substantial questions about boundary conditions.
Encoding precision is not the same as durable learning
A more precise working-memory representation at one moment is not automatically a stronger long-term memory a week later. Working memory, long-term consolidation, later retrieval and transfer are distinct problems.
This distinction protects students from a seductive mistake. An unusual example may seize processing resources. A surprising demonstration may be vividly remembered. Yet the learner can remember the novelty while failing to learn the rule the novelty was supposed to illustrate.
Educational usefulness depends on what was encoded, what relationships were formed, and whether the target knowledge can later be retrieved without the original surprise.
Illustrative case: the strange chemistry example
The following case is illustrative, not research data.
Maren studies acids and bases. Her tutor gives an unusual real-world example involving a substance she has never encountered. The example is distinctive, and Maren remembers it the next day. She concludes that the lesson worked brilliantly.
Then the tutor removes the unusual story and presents three ordinary chemical scenarios. Maren can describe the story but cannot classify the reactions reliably.
The distinctive event survived. The conceptual decision rule did not.
The repair is not to remove all interesting examples. It is to force a second representation: What feature made the example an acid-base case? Which feature was merely memorable? What new case would preserve the rule while changing the surface?
Now the surprise becomes an entry point rather than the memory target.
Prediction can expose what context is carrying for you
One educationally useful implication of predictive processing is that learners often benefit from noticing when an answer was reconstructable from context rather than independently retrievable.
Consider a vocabulary sentence where the surrounding words almost give away the target definition. Correctly completing the sentence can look like strong vocabulary knowledge. Remove the sentence and ask for a definition, synonym, contrast or fresh usage, and performance may change.
The same issue appears in mathematics. A page titled “Pythagoras’ Theorem” makes method selection easy. A mixed problem set removes that contextual prediction. A student who succeeds only when the chapter label supplies the method has not yet demonstrated independent discrimination.
Strategic Resource Allocation does not prove this classroom phenomenon. But it sharpens a useful diagnostic distinction: information can be accessible because its representation is strong or because the surrounding context makes reconstruction cheap.
The reconstruction test
Students can test whether context is carrying too much of the load with four passes.
Pass 1 — Learn in context
Read the explanation, example or worked solution normally. Context is not the enemy; it is how much learning begins.
Pass 2 — Remove the predictive frame
Hide the heading, nearby formula, worked example or sentence context. Ask for the target relation from a neutral cue.
Pass 3 — Change the surface
Use a new example whose superficial details do not point toward the answer. If performance collapses, the original context may have been doing more reconstruction than the learner realised.
Pass 4 — Return after delay
Test again when the original wording is no longer fresh. Durable learning has to survive without the exact predictive sequence that accompanied encoding.
When predictability helps rather than hurts
Predictability is not a flaw. Good instruction deliberately builds predictable structure. A novice benefits when notation is stable, examples are coherent and explanations follow a comprehensible sequence. Predictability reduces unnecessary search.
The problem begins when fluency inside that structure is mistaken for independent capability outside it.
The correct instructional move is not constant novelty. It is staged removal of support. First make the pattern clear. Then vary what should not matter. Finally require the learner to identify the invariant when context stops announcing it.
Why “make it surprising” is a dangerous shortcut
Novelty can capture attention without serving the learning objective. A teacher can create a highly memorable demonstration that leaves students remembering the explosion, anecdote or joke but not the causal mechanism.
Even when surprise genuinely increases encoding effort, there are at least four unanswered questions: What exactly received the extra processing? Did the learner connect it to prior knowledge? Did it persist into long-term memory? Did it transfer to a less surprising case?
Without those checks, “engagement” can become a story told about attention rather than evidence of learning.
Delayed and independent performance check
After a lesson containing distinctive or unexpected material, wait. Then test the target idea with an ordinary example and neutral wording.
Ask three questions separately: Can the learner recall the unusual event? Can the learner explain the target principle? Can the learner use the principle in a fresh case that lacks the memorable surface?
Those answers can diverge. That divergence is informative. It shows whether the distinctive encoding attached to the educational invariant or merely to the episode.
A second check can reverse the direction: give a predictable example but ask an unexpected question about it. If the learner must reorganise the same knowledge rather than recognise a familiar prompt, the assessment becomes less dependent on the original predictive frame.
For parents: memorable is not the same as learned
Children often return from a lesson talking about the most unusual thing that happened. That is valuable evidence of attention and experience, but not yet proof that the curriculum target survived.
A useful follow-up is: “What did that example show?” Then ask for a second example with different surface features. If the child can identify the same relation again, the memorable event may have been integrated into a broader model.
If the child can only retell the event, do not dismiss the lesson. Use the retained episode as a retrieval cue and build the concept outward from it.
For tutors and teachers: manage prediction, then remove it
Instruction can use prediction deliberately without turning novelty into theatre. Ask learners to predict the next step in a proof, the next word in a sentence, the direction of a scientific effect or the likely consequence in a historical argument. When the prediction fails, pause long enough to identify which assumption failed.
Then protect against episodic overlearning. Re-present the principle in a quieter case. Change the wording. Change the layout. Mix the problem among alternatives. Delay the test.
The broader eduKateSG guide HSW-0225 on mixed-set working memory owns the question of how different item sets interact with temporary capacity. HSW-0242 owns a different problem: whether a limited memory system allocates encoding precision strategically according to predictability, and how cautiously that theory should be translated into studying.
Misconceptions
“Unexpected information is always remembered better.” The 2026 paper does not establish that universal educational claim. It proposes and tests an efficiency principle in language processing.
“Surprisal means emotional surprise.” No. In the paper, surprisal is a contextual predictability measure.
“Harder processing means better learning.” Not necessarily. Processing difficulty can arise for many reasons, including confusion and overload.
“Predictable lessons are bad.” Predictability can reduce extraneous search and support novices. The key question is whether capability survives after predictive support is removed.
“If context helps recall, the memory is fake.” Context-supported reconstruction is a normal part of cognition. The issue is whether the learner also needs portable access under the conditions that matter.
Evidence and limits
The primary source is Xu and Futrell, “Strategic resource allocation in memory encoding: An efficiency principle shaping language processing”, published in the February 2026 volume of the Journal of Memory and Language.
The paper develops a computational resource-rational account and examines its predictions with naturalistic production corpora and reading-time data. The evidence is indirect with respect to encoding precision: the authors infer support from patterns in dependency locality rather than directly measuring a memory trace in a classroom learner. Cross-linguistic variability is substantial, and some effects differ by dependency relation and language structure.
Accordingly, this article treats the classroom protocols as proposed applications to be validated through independent performance. The source supports a theory of selective encoding in language processing; it does not prove that adding surprise to lessons improves long-term educational outcomes.
The return
A limited memory system should not be expected to spend equally on everything.
The interesting possibility raised by Strategic Resource Allocation is that prediction helps decide where precision is most valuable. What the context can cheaply reconstruct may need less protection. What the context could not have guessed may deserve more.
For studying, the discipline is to resist turning that elegant idea into a gimmick. Do not chase surprise. Test representation.
After the colourful example, remove the colour. After the predictable chapter, remove the heading. After the explanation, remove the explanation. Then see what the learner can still retrieve, discriminate and use.
That is where an encoding hypothesis becomes an educational question worth answering.
Continue through the How Studying Works Numbered Series Reading Index, or return to the How X Works Hub.