Teaching an AI Tutor to Account for an Individual Student's Recurring Mistakes

A student misapplies a rule several times. The tutor stores a lasting weakness and keeps giving easier exercises even after the learner understands. Such personalization can trap someone in an outdated view of their preparation. Difficulty records need specific answers, boundaries, and regular revision. A mistake informs the next practice choice rather than defining a personality trait to preserve forever.

Here, ‘teaching the tutor’ means configuring instructions, materials, and profile use, not necessarily technically fine-tuning model weights. Memory and context capabilities vary by service and need separate checks. The pedagogical basis is to identify demonstrated performance, form a cautious difficulty hypothesis, and test it through a new action before changing help, preserving the route to independent knowledge application.

Describe the Mistake as an Observation

‘Does not understand fractions’ is too broad. ‘In two equal-numerator tasks, chose the fraction with the larger denominator and explained it by the larger number’ is more useful. It records action, conditions, and reasoning that can be checked and refined. It supports targeted practice instead of repeating a whole chapter without knowing what needs attention or what is already correct.

Record correct actions alongside errors. Correctly comparing equal-denominator fractions is an important boundary. A profile should reveal distinctions rather than only deficiencies, or the assistant may waste time explaining mastered material. Record observed conditions and areas lacking evidence. Personalization remains proportional to actual work rather than an impression from several unsuccessful answers.

Mark assistance levels separately. A correct answer after a complete hint differs from an independent solution, though it can reveal partial understanding. Record exactly what needed prompting. An unconditional average score hides independence changes and fails to distinguish learning from success dependent on constant tutor support.

Do Not Conclude from One Case

An incorrect result may reflect a typo, ambiguous instructions, an unfamiliar word, or a subject gap. Request the reasoning and give another short case. Repeated causes strengthen the hypothesis; otherwise, do not label a recurring error. Check the task too: poor wording rather than misunderstanding may explain a different reading.

Use clear statuses: observed, being checked, confirmed under specified conditions, and no longer appearing in control attempts. Avoid precise confidence percentages without a justified method. A small course needs brief evidence and the next question. This prevents an earlier mistake becoming permanent truth and lets teachers inspect why difficulty or assistance changed.

Separate learning-action profiles from psychological conclusions. Character, motivation, or supposedly fixed learning styles are unnecessary explanations. A subject description suffices for targeted practice. Test recognition of a missed condition or practice a misapplied rule. These are more precise and verifiable than broad labels an assistant can state persuasively without grounds about a person's real circumstances.

Build a Small Map of Component Skills

Break the objective into required actions. Fraction comparison may need understanding parts of a whole, numerator and denominator meanings, and selecting a comparison method. Do not model the entire subject immediately. Identify the prerequisite explaining the particular difficulty. CMU's learning principles help connect component skills and their integration. In personalized practice, the map guides a next task checking the relevant knowledge component.

Prepare correct-action examples and typical incorrect explanations. The tutor must distinguish substance. Looking only at final numbers can miss faulty logic that coincidentally succeeds. Request brief reasoning or an intermediate step. Evidence should suffice without turning every simple exercise into a distracting lengthy report.

Define a new control case for each skill, preserving its foundation while changing surface details. Otherwise, a familiar answer may be memorized. The map should test transfer. If a new exercise reveals another difficulty type, revise the hypothesis rather than force every error into an old profile entry because it simplifies the progress narrative.

Use Lessons from Knowledge Models Without Overgeneralizing

Corbett and Anderson's 1994 work described knowledge tracing in a particular rule-model programming tutor. This shows how action histories can assess knowledge states, but does not imply every modern chat has an equally validated student model. Distinguish research architecture, an actual implementation, and a simple text profile supplied to a generative assistant.

An initial scenario needs no complex probabilistic system. Start with a short card: skill, evidence, hypothesis, next check, and date. Update it with new data. Automatic adaptation across a large program needs appropriate design and evaluation specialists. A full chat history is not a complete diagnostic profile merely because it contains every earlier exchange.

Check which profile information the tool actually receives. Context can be limited, shortened, or changed by settings. Storing a record does not guarantee correct use. Supply a known profile and new exercise, then check whether help matches the need. If the current card is ignored or an old one used, change the process before student use rather than merely adding more conversational data.

Organize Assistance Around the Hypothesis

For a confirmed difficulty, use an exercise requiring the relevant distinction. With fractions, keep the same whole and numerator while varying denominators. The student explains changing part size, then tries another example. This targets a missing distinction rather than merely adds topic exercises. The tutor should check whether the distinction appears independently.

Do not repeat one hint endlessly. If students repeat it without understanding, change representation or return to a prerequisite. Compare a diagram, verbal explanation, and short counterexample where appropriate and accessible. Do not declare one their permanent learning style. Choose by the required action and evidence of learning rather than an unchecked profile label.

After assistance, provide a new attempt without the same hint. One success should not automatically remove every limitation. Define sufficient repeated checking for the task. If mistakes disappear under different relevant conditions, update status and difficulty. The profile should not keep learners on easy tasks or explain what they can now perform and justify independently.

Let the Student Clarify the Profile

Show the record in ordinary language. Participants should understand which difficulties are considered and why. They may explain a typo or demonstrate a new correct method. Do not automatically treat every self-description as evidence; verify the action. Without this route, the profile becomes an opaque decision students cannot understand or discuss meaningfully with a teacher.

Provide a short control case or human review for disagreement rather than exhausting argument with the assistant. Resolve the issue through subject evidence. Record why the profile changed so later assistance does not restore outdated assumptions. This matters when several teachers or tools interpret the same learning path's attempts and independence differently.

Avoid stigmatizing wording. ‘Needs a denominator-choice check under these conditions’ describes an action. ‘Weak at mathematics’ hides details and can shape expectations. UNESCO addresses human responsibility in GenAI use. Here, that includes reviewing automated assumptions, explaining their grounds, and restoring access to more challenging practice after demonstrated progress.

Collect Only Necessary Data

Choosing the next exercise often needs only the skill, attempt conditions, assistance type, and check result. A full personal profile, private stories, or unrelated assessments may be unnecessary. State each field's purpose and defer collection if it does not affect a teaching decision. Memory volume is not personalization's goal. A small current record correctly linked to the next task and accessible to responsible participants matters more.

Check storage and access in the actual tool. Do not assume every text profile may be transferred externally. Use fictional cards for a pilot. NIST Privacy Framework provides a broader privacy-risk foundation. Define local purpose, necessary data, and responsibilities rather than treating ‘personal tutor’ as proof of suitable student-data conditions.

Define currency periods and review triggers. After a module, some entries no longer serve the current route; explain why they are retained if applicable. Do not transfer old difficulties automatically to another subject. A familiar term may signify another action. The profile should support learning in context rather than accumulate a comprehensive error history for undefined future personal decisions.

Test Adaptation with Control Cards

Before launch, create fictional profiles: a confirmed recurring error, one doubtful case, a mastered skill, and conflicting data. Define the expected next action for each. The tutor should seek clarification when grounds are weak and stop basic explanations after confirmed progress. This checks profile-based assistance logic rather than conversational pleasantness.

Include a correct answer in unusual wording. Rejection for differing from a template can create false profile entries. Also test an incorrect answer with persuasive reasoning. Smoothness must not hide substance. NIST GenAI Profile helps treat unreliable outputs as a risk. Your pilot needs control cases showing their consequences for subsequent student practice.

After instruction changes, repeat decisive tests with new examples. Passing an old control answer alone is insufficient. Check changed logic rather than familiar-text memorization. Save conditions and a short result so testing can be repeated after service updates, profile changes, or new modules without relying on one successful demonstration.

A Hypothetical Learning-Hypothesis Update

A student repeatedly compares equal-numerator fractions by denominator size. The initial card records observation and hypothesis rather than a general subject judgment. The tutor requests reasoning and another same-whole case. After targeted analysis, the student independently compares a new pair and explains part size. This illustrates updating rather than reports a measured system or product effect.

The next check changes numbers and representation. If reasoning survives without hints, update the card: earlier difficulties did not appear under control conditions, so progress is possible. A later mistake need not have the same cause. Read its explanation again: different wholes or presentation may now matter. Keep the profile responsive rather than let an old hypothesis absorb every new observation.

The teacher checks whether assistance actually changed. Continued explanation of mastered foundations means personalization is wrong even if records are technically stored. The useful result is suitable new practice with sufficient challenge. Students should see evidence change the path, enabling discussion through concrete materials rather than an opaque ability ranking unrecoverable from individual actions.

Prioritize When There Are Several Mistakes

A profile can contain different difficulties without the next exercise fixing all simultaneously. Identify the prerequisite blocking the current action and start there, keeping other observations visible for later. This reduces load and clarifies change causes. A remaining error after foundational learning needs its own targeted task rather than another repetition under a broader topic label.

Ask the tutor to explain exercise choice through a particular record: checking part-size understanding before comparing harder fractions, for example. Teachers can assess and revise the logic; students see the practice's purpose. Every internal detail need not be displayed, but decisive choices should be understandable. Addressing a learner by name does not make an exercise personalized if it lacks current evidence.

Distinguish a Returning Error from a New Context

After a break, a student may forget a rule or fail to recognize it in another representation. The first needs recall, the second condition comparison. Request a short explanation and a control example rather than automatically restoring the old hypothesis. Consider changes in time, format, subject conditions, or prior assistance.

If a student explains a rule correctly but fails in an unfamiliar situation, avoid unnecessary lengthy definitions. Compare two cases and ask for their shared foundation, then change surface details. Check the result with another attempt. A different assistant explanation alone does not establish that the decisive difficulty or its cause has been resolved.

Check Whether the Profile Itself Helps

In an internal pilot, compare exercises chosen with and without a current card on the same control case. This does not prove learning effects, but reveals whether the record changes help appropriately. Without a difference, check format, volume, and instruction clarity. A small structured card with a specific next question may work better than an unprioritized, undated conversational history.

Also test an outdated card needing revision after a correct answer. If the assistant clings to it despite evidence, change update rules or leave the decision to the teacher. Personalization must correct its own assumptions. Otherwise, an initial misjudgment repeats across assignments and wastes practice time on disproved or overly broad hypotheses.

Connect the Profile to the Learning Path

The WeStudy LMS platform can organize materials and practice sequence. Check storage and transfer of a particular profile to an external assistant separately. Do not claim automatic diagnosis without a verified implementation. Start with a clear card used by a teacher to select tasks. Difficulty records need subject evidence, a date, and a next check whose result actually affects subsequent activity.

When the WeStudy LMS platform is part of a program, review profiles alongside new independent attempts. Good personalization reduces unnecessary assistance after mastery and restores it when evidence shows a new need. It must not trap people in old descriptions. Quality appears in appropriate practice, explainable progress, and increasingly independent performance under new conditions.

Before the next module, check that current records help select the required action rather than obscure new subject needs.

Sources

• Corbett, Anderson: Knowledge tracing — Modeling the acquisition of procedural knowledge.

• Carnegie Mellon University: Learning Principles.

• UNESCO: Guidance for generative AI in education and research.

• NIST: Generative Artificial Intelligence Profile.

• NIST: Privacy Framework.