How Machines Adapt to Individual Learners

In 1984 Benjamin Bloom published a finding so striking that it haunted education researchers for four decades: students tutored one-to-one outperformed classroom-taught peers by two standard deviations, placing the average tutored student above 98 percent of the comparison group.

The Promise That Started Everything

In 1984, educational psychologist Benjamin Bloom presented a paper that would become the most cited provocation in the history of educational technology [1]. Drawing on dissertation studies by Anania and Burke, Bloom reported that students who received one-to-one tutoring combined with mastery learning outperformed students in conventional classrooms by roughly two standard deviations. The number was staggering. It meant the average tutored student performed better than about 98 percent of peers taught in a regular classroom [2]. Bloom called it the "two sigma problem" and framed it as a challenge: can we find methods of group instruction as effective as one-to-one tutoring? Because universal one-to-one tutoring, he acknowledged, was "too costly for most societies to bear on a large scale."

That challenge launched a research program that spans half a century, from mechanical boxes on teachers' desks to neural networks trained on hundreds of millions of learning interactions. The question "can machines adapt to individual learners" is not futuristic. It is the oldest question in educational technology. And the answer, as of 2026, is more nuanced and more interesting than either enthusiasts or skeptics tend to admit.

A recent reassessment published in Education Next revealed something uncomfortable about Bloom's famous figure: the graph reproduced in thousands of papers and presentations was hand-drawn in a "smooth, stylized fashion to show what a two-sigma effect might look like" rather than plotted from actual data [3]. The two-sigma benchmark, it turns out, was always partly aspirational. But even if the real advantage of human tutoring is closer to one sigma than two, the gap between mass instruction and individualized attention remains enormous. And the question remains urgent: can machines close it?

The Man Who Built a Box to Replace a Teacher

The idea that a machine could adapt to a learner did not begin with computers. It began with a psychologist, a piece of plywood, and a conviction that classrooms were broken.

B.F. Skinner, the Harvard behaviorist best known for training pigeons, published "The Science of Learning and the Art of Teaching" in 1954. His frustration was specific: a classroom of thirty students meant that individual reinforcement, the kind that shapes behavior effectively, was diluted to nearly nothing [4]. A teacher asking a question to a room gets one response from one student. The other twenty-nine sit idle. Skinner's solution was a mechanical device that broke material into small "frames," required an active written response to each frame, provided immediate feedback on correctness, and allowed each student to proceed at their own pace. That is the original adaptive learning loop. Present, prompt, respond, reinforce.

Skinnerian programmed instruction swept through American schools in the early 1960s. By 1962, teaching machines were a national conversation. By 1970, they were gone [5]. The content was boring. The branching was rigid. Teachers resented being replaced by boxes. And the theoretical framework, strict operant conditioning, turned out to be too narrow for the messy reality of human learning [6].

But two things survived. First, the principle that individualized pacing matters. Second, the architectural pattern: small steps, active responding, immediate feedback, self-paced progression. Every adaptive system built since, from the simplest flashcard app to the most complex neural-network tutor, inherits that architecture.

When Machines Learned to Think About Thinking

The leap from Skinner's mechanical frames to genuine intelligence happened in 1970, when Jaime Carbonell at Bolt, Beranek and Newman built SCHOLAR [7]. SCHOLAR tutored South American geography, but its real innovation was not the subject matter. It was the architecture. Instead of following a pre-scripted decision tree, SCHOLAR stored knowledge as a semantic network and generated questions dynamically based on what the student had and had not yet demonstrated knowing [8]. For the first time, a machine was reasoning about a learner's knowledge state, not just advancing through a fixed sequence.

The 1970s and early 1980s produced a burst of systems that pushed this idea further. Allan Collins and Albert Stevens built WHY in 1977, a Socratic tutor for meteorology that could present cases, ask for predictions, probe for missing factors, entrap the student when they had not considered all variables, and generate counterexamples [9]. SOPHIE, built by John Seely Brown, Richard Burton, and Johan de Kleer for U.S. Navy electronics training, used a working circuit simulation as its knowledge base, allowing students to insert faults and reason through their consequences in natural-language dialogue [10].

These early intelligent tutoring systems (ITS) shared a structure that still defines the field: a domain model (what the system knows about the subject), a student model (what the system believes the learner knows), and a pedagogical model (how the system decides what to do next). The student model was the critical innovation. For the first time, the machine maintained an internal representation of an individual learner's knowledge, updated it after every interaction, and used it to make decisions.

But the most consequential research program came from John R. Anderson at Carnegie Mellon University. Anderson's ACT-R cognitive architecture modeled human procedural skill as a collection of production rules, each associated with specific brain regions [11]. His group built tutors that traced a student's solution path against an expert model, step by step, intervening precisely when the student's reasoning diverged. Anderson, Boyle, and Reiser's 1985 paper in Science introduced the LISP Tutor and the Geometry Tutor and declared that "cognitive psychology, artificial intelligence, and computer technology have advanced to the point where it is feasible to build computer systems that are as effective as intelligent human tutors" [12].

That claim was premature. But the model-tracing approach worked well enough to become Carnegie Learning's Cognitive Tutor, later renamed MATHia, which has been deployed in thousands of American classrooms and remains one of the few adaptive systems with a serious body of randomized controlled trial evidence.

How Machines Actually Track What You Know

The engine behind modern adaptive learning is the student model. And the two most influential approaches to student modeling both emerged from Carnegie Mellon.

In 1995, Albert Corbett and John Anderson published "Knowledge Tracing," introducing Bayesian Knowledge Tracing (BKT) [13]. BKT is a two-state hidden Markov model. At any moment, a student either has or has not mastered a particular skill. The system cannot observe mastery directly; it can only observe correct or incorrect responses, which are noisy indicators because students sometimes guess correctly without knowing (the "guess" parameter) and sometimes make mistakes despite knowing (the "slip" parameter). Four numbers define the model for each skill: prior knowledge probability, learning rate, guess rate, and slip rate. After each student response, the system updates its estimate of mastery using Bayes' theorem [14].

BKT was the workhorse of the Cognitive Tutor for two decades. Its elegance lies in interpretability: every parameter has a clear meaning, and the system's belief about a student is always a single probability between zero and one.

Twenty years later, in 2015, a Stanford team led by Chris Piech proposed Deep Knowledge Tracing (DKT) at the NeurIPS conference [15]. DKT replaced BKT's hand-crafted skill encodings with a recurrent neural network (an LSTM) that learns latent representations of student knowledge directly from sequences of interaction data. No one needs to define what "skills" exist. The network figures it out from patterns in the data. On several public benchmarks, DKT substantially outperformed BKT [16].

The tradeoff is transparency. BKT tells you "this student has a 73 percent probability of having mastered single-digit addition." DKT tells you "given this student's entire response history, the probability of a correct answer on the next problem is 73 percent." The prediction may be more accurate. But the underlying knowledge state is a vector of numbers that no human can interpret.

A parallel tradition, Item Response Theory (IRT), comes from psychometrics rather than AI. IRT models estimate a student's latent ability on a continuous scale and each item's difficulty, discrimination, and guessing parameters. Computerized Adaptive Testing (CAT), used in the GRE and many standardized exams, dynamically selects questions to maximize information about a test-taker's ability, typically reducing test length by about 50 percent without loss of measurement precision [17] [18].

And then there is Knowledge Space Theory (KST), the mathematical framework behind ALEKS (Assessment and Learning in Knowledge Spaces). Developed by Jean-Claude Falmagne and Jean-Paul Doignon in 1985, KST models a domain as a partially ordered set of knowledge states. Algebra, for instance, is modeled as roughly 350 concepts that give rise to millions of feasible knowledge states. A Bayesian diagnostic engine identifies a student's current state in about 20 to 25 questions [19].

The Purest Form of Machine Adaptation

There is one class of adaptive system where the adaptation is mathematically precise, individually calibrated, fully transparent, and backed by over a century of empirical evidence. It is also the humblest: spaced repetition.

A spaced repetition algorithm estimates, for each combination of learner and item, an explicit forgetting trajectory, and schedules the next review to occur just before predicted recall failure. Unlike most "AI tutors," which adapt through opaque embeddings that no one can inspect, a spaced repetition scheduler shows you exactly what it believes about your memory and why.

The lineage starts with Hermann Ebbinghaus in 1885 and his exponential forgetting curve. Sebastian Leitner turned the principle into a physical system of cardboard boxes in 1972. Piotr Wo?niak created the first computational scheduling algorithm, SM-2, in 1987 in Pozna?, Poland. SM-2 assigned each card an "ease factor" that adjusted based on self-rated difficulty. It was simple, documented, and effective enough to become the default engine for most digital flashcard software built in the following two decades.

In 2016, computational linguists Burr Settles and Brendan Meeder published a paper at ACL describing Half-Life Regression (HLR), trained on 13 million user-word practice traces from a language learning platform [20]. HLR estimates each vocabulary item's memory half-life as a function of the learner's practice history, and Settles reported that its prediction error was nearly half that of a Leitner-style baseline [21] [22].

The most recent breakthrough is FSRS, the Free Spaced Repetition Scheduler. In August 2022, Jarrett Ye published a paper at ACM KDD proposing a new scheduling approach using stochastic dynamic programming. FSRS fits a three-component memory model (stability, difficulty, retrievability) to each user's individual review history [23]. Since late 2023, FSRS has become the default scheduler in the largest open-source flashcard platform. Per the Expertium benchmark on roughly 700 million anonymized reviews, FSRS-5 achieves a log-loss of 0.291 and RMSE of plus or minus 5.3 percent, compared to SM-2's log-loss of 0.354 and RMSE of plus or minus 16.2 percent [24]. That translates to roughly 25 percent fewer daily reviews for the same 90 percent retention target.

Why does spaced repetition matter for the broader question of machine adaptation? Because it is the existence proof. It demonstrates that machines can track individual learners with mathematical precision, adapt to their specific forgetting patterns, improve with every interaction, and do all of this transparently. The science behind how spaced repetition works is among the most replicated in all of cognitive psychology. And the evolution from SM-2 to FSRS shows how machine-learning methods can push individual adaptation further than hand-crafted rules ever could.

What Brains Actually Differ In (And What They Do Not)

If machines are to adapt to individual learners, they need to know what varies between individuals. And the most popular answer to this question, the one embedded in corporate training platforms and teacher preparation textbooks alike, is wrong.

The notion that people have distinct "learning styles," visual or auditory or kinesthetic, and that matching instruction to a student's preferred style improves outcomes, is one of the most persistent myths in education. In 2008, Pashler, McDaniel, Rohrer, and Bjork published a commissioned review in Psychological Science in the Public Interest that examined the evidence rigorously [25]. Their conclusion was blunt: "at present, there is no adequate evidence base to justify incorporating learning-styles assessments into general educational practice" [26]. The required experimental evidence, a crossover interaction where visual learners perform better with visual instruction AND auditory learners perform better with auditory instruction, essentially never appears.

And yet the myth persists. Nancekivell, Shah, and Gelman found in 2020 that more than 90 percent of their mixed sample endorsed the learning-styles hypothesis [27] [28]. Educators of young children were most likely to hold "essentialist" beliefs that learning style is innate. A 2023 review in Frontiers in Education confirmed the persistence of this "neuromyth" even among trained teachers [29].

So what does actually differ between learners? The neuroscience points to several dimensions that matter enormously and that good adaptive systems can target.

First, working memory capacity. Alan Baddeley's tripartite model, published in Nature Reviews Neuroscience in 2003, describes a central executive that coordinates a phonological loop (for verbal information) and a visuospatial sketchpad (for spatial information), plus an episodic buffer that integrates them [30]. Individuals differ markedly in how much they can hold and manipulate simultaneously. Randall Engle's research on executive attention showed that these differences are rooted in the dorsolateral prefrontal cortex and predict both fluid intelligence and learning rate across domains [31].

Second, learning rate heterogeneity at the neural level. In 2024, Muller and colleagues published a striking finding in Nature Neuroscience: different prefrontal neurons encode systematically different learning rates from positive versus negative prediction errors [32]. This "distributional reinforcement learning" means the brain does not have a single learning rate. It has a population of neurons, each updating at its own speed, whose aggregate produces an individual's characteristic rate of adaptation.

Third, the hippocampal-prefrontal axis and memory consolidation. Miller and Constantinidis argued in a 2024 review in Nature Reviews Neuroscience that the prefrontal cortex orchestrates short-term and long-term memory systems on multiple timescales simultaneously [33]. The rate at which new learning is consolidated from hippocampal short-term storage to neocortical long-term storage varies between individuals and depends on factors including sleep quality, prior knowledge structures, and stress.

The practical implication is clear. Good adaptive systems should adjust to what a learner knows (prior knowledge), how fast they learn (learning rate), how much cognitive load they can handle (working memory capacity), and how quickly they forget (forgetting trajectory). These are real, measurable, neurally grounded variables. They are nothing like "visual versus auditory."

Do They Actually Work? The Evidence

The question "can machines adapt" is interesting. The question "does it help students learn" is the one that matters.

The most rigorous answer comes from four major meta-analyses. Kurt VanLehn's 2011 review in Educational Psychologist compared human tutoring, step-based intelligent tutoring systems, and answer-based computer-assisted instruction [34]. Human tutoring produced effect sizes of approximately d = 0.79. Step-based ITS, the kind that traces student reasoning step by step, came close. Answer-based CAI, which only evaluates final answers, was substantially weaker at d = 0.31. The headline: well-designed intelligent tutoring systems approach, though do not exceed, the effectiveness of human tutors.

Ma, Adesope, Nesbit, and Liu (2014) conducted a meta-analysis of ITS outcomes and found positive effects over conventional instruction with an effect size of approximately g = 0.41 [35]. Steenbergen-Hu and Cooper (2014) found small but positive effects for K-12 mathematics, with an important caveat: effects were smaller for low-achieving students [36].

The largest and most recent meta-analysis, by Kulik and Fletcher in 2016, analyzed 50 controlled evaluations and found a median effect size of d = 0.66. Students receiving intelligent tutoring outperformed conventionally instructed students in 92 percent of the evaluations [37] [38]. A separate evaluation of the DARPA Digital Tutor for U.S. Navy IT training reported effect sizes as high as d = 1.97, approaching Bloom's benchmark, though this reflected an extraordinarily intensive implementation that is difficult to generalize.

System-specific evidence adds texture. A large-scale RAND randomized controlled trial of Carnegie Learning's Cognitive Tutor found modest but statistically significant gains of about 0.20 standard deviations for high school algebra over two years [39] [40]. A meta-analysis of ALEKS by Fang, Ren, Hu, and Graesser found the system "as good, but not better than, traditional classroom teaching" [41]. DreamBox Learning showed a 0.12 standard deviation gain in a larger randomized trial [42].

In China, Squirrel AI reported striking gains in internal studies [43] [44], though independent peer-reviewed replications remain sparse.

The honest summary: adaptive systems produce medium-sized, replicable effects. They work. They do not yet deliver on Bloom's two-sigma promise. And the gap between what marketing departments claim and what randomized trials show remains wide.

The GPT-4 Tutors: A Step Change and a Warning

In 2023, the field shifted. Large language models, specifically GPT-4, made it possible to build tutoring systems that could hold sustained, mixed-initiative dialogue with students, generate explanations on the fly, ask Socratic questions, and respond to natural language in ways that earlier ITS could not.

Khan Academy launched Khanmigo in March 2023, built on GPT-4 and designed with explicit guardrails against simply giving answers [45]. Sal Khan described its philosophy: "Unlike other AI tools, Khanmigo doesn't just give answers. It guides learners to find the answer themselves" [46]. By the 2024-2025 school year, Khanmigo had expanded to over 700,000 student and teacher users across more than 380 district partners [47].

The first strong peer-reviewed evidence arrived in 2025. Kestin, Miller, Klales, Milbourne, and Ponti at Harvard built PS2 Pal, a GPT-4-based tutor for introductory physics, and tested it in a within-subjects crossover design with 194 undergraduates. Students "learned more than twice as much in less time" using PS2 Pal compared to research-based in-class active learning [48]. This is impressive. It is also a single course at a highly selective institution with intensive instructor scaffolding.

The most sobering evidence came from the same year. Hamsa Bastani, Osbert Bastani, Alp Sungu, and colleagues at the University of Pennsylvania and Wharton conducted a pre-registered randomized controlled trial with approximately 1,000 Turkish high-school math students [49]. Students were randomly assigned to one of three conditions: GPT-4 Base (unmodified ChatGPT access during practice), GPT-4 Tutor (a pedagogically scaffolded version that withheld answers and guided through hints), or a control group with no AI access.

The results were striking. During practice, GPT-4 Base students solved 48 percent more problems correctly. GPT-4 Tutor students solved 127 percent more. The AI was clearly helping in the moment. But then the researchers removed AI access and gave an unassisted exam. GPT-4 Base students scored 17 percent worse than controls [50]. The easy availability of instant answers had short-circuited the productive struggle that drives durable encoding. GPT-4 Tutor's guardrails mostly eliminated this harm, but the finding was clear: unguarded AI access does not just fail to help. It actively damages learning.

A follow-up study in 2026 using chess clubs produced a parallel finding [51]. Students with on-demand AI assistance achieved less than half the performance gains of those whose AI access was rate-limited (30 percent improvement versus 64 percent). Even high-skill players over-requested help when access was unrestricted. As Bastani put it: "Using AI as a tutor is like keeping a big jar of cookies in the kitchen cabinet. You tell yourself you're just going to eat one, but it's a slippery slope. Self-regulation is hard, even when you know something isn't good for you."

The concept maps directly to Lev Vygotsky's Zone of Proximal Development: the band of tasks a learner cannot yet do alone but can do with appropriate support. When AI makes tasks too easy, it pushes the learner outside the ZPD. The task no longer requires the cognitive effort that produces learning.

The Risks That Nobody Talks About

Beyond over-reliance, several underappreciated risks attend the deployment of adaptive learning systems at scale.

The cold-start problem is inherent to every personalized system. A learner model begins with zero data. For a student's first session, the system knows nothing and must fall back on population averages. Some systems use explicit pretests to accelerate cold-start diagnosis. Others use meta-learning to transfer parameters from similar learners. But the first few interactions are always the least adaptive.

Algorithmic bias is a structural concern. If a knowledge-tracing model is trained predominantly on data from suburban American middle-school students, its "guess" and "slip" parameters may systematically misfit students whose cultural background, language, or test-taking conventions differ. Steenbergen-Hu and Cooper explicitly noted that ITS effects were smaller for low-achieving students, a result that could reflect bias in training data or curriculum design.

UNESCO's 2023 guidance on generative AI in education flagged algorithmic bias as a primary governance concern and recommended a minimum age of 13 for independent use of generative AI tools [52] [53].

Privacy is the flip side of personalization. Adaptive systems require fine-grained behavioral data: response latencies, error patterns, hint requests, navigation paths. The same data that enables individualized instruction enables surveillance. Federated learning, where models are trained on local devices and only aggregated gradients are shared, offers a technical mitigation. But institutional and regulatory frameworks have not kept pace with the technology.

And there is a subtler risk that receives almost no attention: filter bubbles in learning. If an adaptive system optimizes for short-term mastery, it may systematically avoid exposing learners to challenging, surprising, or interdisciplinary material. The pedagogical equivalent of "engagement maximization" could produce students who are locally fluent but globally narrow, who master the prescribed curriculum but never encounter the unexpected idea that sparks genuine intellectual growth.

Where This Goes Next

The frontier of adaptive learning in 2026 involves several directions that the current generation of systems barely touches.

Neuroadaptive learning aims to close the loop with physiological signals. Consumer-grade EEG headsets, eye-tracking cameras, and pupillometry can estimate cognitive load, attention, and frustration in real time. The technical challenge is noise. The ethical challenge is surveillance. The pedagogical promise is real: imagine a tutoring system that detects when working memory is overloaded and automatically reduces problem complexity before the learner becomes frustrated.

Affective computing, associated with researchers like Arthur Graesser and Sidney D'Mello, seeks to detect confusion, boredom, and engagement from interaction patterns and to respond accordingly. The evidence that detecting affect can improve learning outcomes is promising but mixed, and the privacy implications for minors are serious.

Open Learner Models, championed by Susan Bull and Judy Kay, propose that the system's representation of the learner should be visible to the learner. Instead of a black box that decides what to show you next, an open model lets you see what the system thinks you know and do not know. This transforms adaptation from a passive experience into a metacognitive tool.

And the newest thread, emerging in 2025 and 2026, is multi-agent and agentic tutoring architectures [54]. These systems combine a formal learner model (BKT or DKT-style) with a generative reasoning engine (an LLM), mediated by an orchestrator that plans pedagogical strategies. The orchestrator ensures that the LLM's responses are grounded in verified curriculum and calibrated to the learner's ZPD. Early prototypes show promise. Whether they can scale to millions of learners without degrading quality is an open question.

What We Know, What We Do Not, and What It Means

Machines can adapt to individual learners. This is no longer a question. Three decades of randomized trials, four major meta-analyses, and billions of learning interactions processed by algorithms from BKT to FSRS have settled it.

What machines adapt to matters more than whether they adapt. The empirical case against "learning styles" is overwhelming. The empirical case for adapting to prior knowledge, working memory load, error patterns, and forgetting trajectories is strong. Any adaptive system that asks you to take a VARK inventory and personalizes on the result is selling a story that the evidence does not support.

How well machines adapt is the honest frontier. The median effect size of d = 0.66 from Kulik and Fletcher represents a meaningful improvement over conventional instruction. It falls short of Bloom's two-sigma benchmark, which itself was always partly aspirational. LLM-based tutors like Khanmigo and PS2 Pal may represent a step change in interaction quality. But the Bastani findings are a sharp warning: without guardrails, the most powerful tutors produce the most dangerous crutch effects.

The wisest strategy for anyone using adaptive learning tools, whether as a student, an educator, or an institution, is to combine them with human instruction, demand evidence on unaided assessments, prefer systems whose internals are transparent, and remember that productive struggle is not a bug to be optimized away. It is the mechanism by which learning happens.

The machines that best adapt to individual learners are not the ones that make learning feel easiest. They are the ones that keep each learner working at the edge of what they can do alone, expanding that edge one interaction at a time. Vygotsky called it the Zone of Proximal Development. Bloom called it the two-sigma problem. The engineers building the next generation of adaptive systems call it the optimization target. By any name, it is the same thing: the narrow band where effort meets growth. Machines that find it for each learner, and hold them there, will earn their place in education. Machines that bypass it will make learning feel productive while making learners weaker.