# Predicting the next question in an adaptive AI tutor | SwaVid Research | SwaVid A reproducible method for converting curriculum structure and current-session learner evidence into the next valid, informative question. Canonical: https://swavid.com/research/how-swavid-chooses-the-next-question Source: https://swavid.com/research/how-swavid-chooses-the-next-question # Predicting the next question. ## Results at a glance ## The next-question problem is a constrained decision problem. ## From curriculum map to next question: the complete prediction pipeline. ## Build the curriculum frontier before choosing a question. ## Score every valid question by the decision it can resolve. ## The online algorithm selects one question in five steps. ## Worked example: two answers can change the next question. ## Evaluate on chapters the selector did not see during training. ## The selector reproduced the reference policy across every held-out group. ## What is established today, and what should be tested next. ## A reference architecture for curriculum-grounded adaptive tutors. ## Four implementation principles other AI tutors can adopt. ## Definitions for implementing the policy. ## Reference configuration used in the reported evaluation. ## What an implementation team needs to know. ## Reduce uncertainty inside a trusted curriculum frontier. ### Start with NCERT evidence ### Govern directed edges ### Update the gap posterior ### Minimise the next-question cost ### Map prerequisite concepts ### Update evidence after each answer ### Construct valid choices ### Choose the lowest-cost probe ### Combine five interpretable costs ### Remove invalid questions before ranking ### Optionally refine the reference ranking ### The surrogate learns ranking, cost, and confidence together. ### Locate the active frontier ### Build the feasible set ### Score every survivor ### Take the argmin ### Update and repeat ### Friction begins one step earlier: force. ### 4,490 decision groups ### 248 labelled edges ### Policy fidelity is the result being measured. ### Curriculum validity ### Policy implementation ### Learner-state calibration ### Learning impact ### Curriculum-specific evidence ### Independent validity controls ### Current-session evidence ### Reproducible decision contract ### Ground candidates in the curriculum ### Update learner state before selection ### Apply hard constraints before ranking ### Earn model complexity against baselines #### Which statement best describes a force? A reproducible method for turning curriculum structure and a learner’s latest answers into the next valid, informative question. Figure 1 Next-question prediction as constrained decision-making: curriculum evidence defines the frontier, session evidence updates learner state, and the policy selects the lowest-cost valid probe. This whitepaper presents a reproducible next-question prediction method for curriculum-grounded AI tutors. The method separates four decisions that are often collapsed into one: which prerequisite concept is active, what the learner’s answers imply, which questions are valid, and which valid question is most informative to ask next. SwaVid implements these decisions with a reviewed NCERT prerequisite graph, Bayesian-style evidence updates, a hard feasibility mask, and a five-component diagnostic cost. On seven chapter-disjoint test chapters, the deployed selector ranked 11,237 candidates in 930 decision groups, chose a reference-optimal question in all 930 groups, and made zero infeasible selections. This is evidence of reference-policy fidelity—not a claim of perfect learner prediction or measured learning gains. The contribution is an inspectable architecture that another adaptive tutor can implement, test, and improve. The reported scale covers the published curriculum graph and the held-out decision benchmark. “Agreement” means agreement with the defined reference policy, not correctness prediction for learners. 187 question-bearing concept links plus 336 separately audited navigation links. Ranked across 930 decision groups drawn from seven chapter-disjoint test chapters. The deployed selector chose a reference-optimal question in every held-out decision group. A structural property of the hard feasibility mask, verified on all held-out groups. A useful next question must satisfy three conditions at once: it must test the right prerequisite, be valid for the learner’s current state, and reduce uncertainty at an acceptable cost. SwaVid solves these conditions in sequence instead of asking a generator to invent a plausible follow-up. Questions, answers, source passages, concept mappings, and prerequisite candidates enter through reviewed curriculum evidence, not student history. A candidate u → v is publishable only when grounded review supports the relationship and the exact source concept has trusted NCERT question evidence. Each answer changes the probability that the active prerequisite is missing. At least two probes are required to resolve it. Only trusted, feasible questions compete. The policy balances decision risk, uncertainty, learner effort, source risk, and edge risk. Curriculum validity, learner-state estimation, candidate eligibility, and question ranking are independently inspectable. A ranking error cannot make an unsupported question eligible. The offline system builds a trusted prerequisite graph. During a session, each answer updates the active concept’s evidence state; the server then constructs a feasible question set and ranks it. Figure 2 shows the information flow, while Table 1 defines the notation used by the selection rule. Reviewed NCERT evidence identifies the concepts a learner may need before beginning the target chapter. The active session maintains a transparent probability that each prerequisite gap is still unresolved. Only trusted questions for the active concept, within the remaining probe budget, enter the feasible set. The policy ranks every feasible question by diagnostic value, uncertainty, effort, source quality, and graph risk. Inputs available at selection time: the published graph, trusted question metadata, current-session answers, the remaining probe budget, and the promoted selection policy. Offline training records are not required. Figure 2 End-to-end information flow. Curriculum relationships are prepared offline; learner-state updates, feasibility checks, and question ranking execute during the diagnostic session. Table 1 Notation used in Sections 3–5. Symbols are introduced at first use in the text; this table is the consolidated reference. A candidate edge is directed: source concept u must be understood before target chapter v . A candidate is publishable only when the grounded curriculum audit supports the relationship and the source concept has trusted NCERT question evidence. A calibrated classifier helps score this operational label; publication rules remain the final authority. Temperature T and the decision threshold are selected on calibration chapters. Because an invented prerequisite can send a learner down the wrong path, the calibration objective assigns false positives twice the cost of false negatives. Figure 3 A published prerequisite edge. The source concept must be secure before the target chapter is taught. Table 2 Production-graph evaluation coverage at the time of the reported run. Table 3 Performance against the operational prerequisite labels on 161 chapter-disjoint test edges. These labels combine grounded curriculum support with exact-concept question evidence. Figure 4 Held-out confusion matrix for the prerequisite-edge classifier at the operating point reported in Appendix B. Recall is complete on this split; six rejected candidates are false positives and remain subject to publication checks. How to interpret this result. The classifier reproduces the current operational labels with 96.27% accuracy; the trusted-source-evidence feature alone reproduces the same confusion matrix. The scientific value therefore lies in the governed publication procedure, not in claiming that a neural model discovered prerequisite truth. The resulting graph contains 187 question-bearing concept prerequisites and 336 separately audited navigation prerequisites. Once the active concept and feasible set are known, selection becomes a ranking problem. Each candidate receives five interpretable costs: decision risk, remaining uncertainty, learner effort, source risk, and prerequisite-edge risk. The minimum total cost defines the reference choice; a promoted surrogate can reproduce or refine that ranking without overriding feasibility. Decision risk is the expected probability of remaining on the wrong side of the gap decision after the answer. Entropy is expected posterior uncertainty; effort is normalized question demand; source and edge risk are one minus their respective confidence scores. The weights are explicit design parameters that can be recalibrated with learner-outcome data. Figure 5 Reference diagnostic cost used to construct and test the ranking policy. Keeping the components explicit makes each next-question decision inspectable. The reported A100 run used AdamW, gradient clipping, and validation-based checkpoint selection. The composite objective rewards correct ordering, numerical cost agreement, probability mass on the best set, a ranking margin, and a small residual. The loop is executed after every answer. It first identifies the unresolved curriculum frontier, then filters and ranks candidates, asks one question, updates the evidence state, and repeats. Let G mean “the active concept is a material gap,” r t be the latest response, and ρ q be question reliability. The update is performed in log-odds space so lower-confidence evidence has proportionally less influence. The published graph fixes which concept is eligible: a missing prerequisite of the target chapter, not whatever a generator finds plausible. Only trusted, in-concept, in-budget probes survive. In the reported dataset, 49,814 candidate actions form 4,490 decision groups. Each feasible question receives five interpretable cost components. Lower cost means the question is safer and more useful for resolving the current decision. The lowest-cost feasible question wins. A low serving temperature converts cost into a sharply concentrated selection probability without weakening the mask. The learner’s answer updates the gap posterior. The loop stops only when at least two probes resolve the concept at the 95% threshold. Figure 6 The five-step inference procedure executed for every next-question choice in a session. This interactive Friction scenario shows how the posterior changes after correct and incorrect evidence, and why one response is not enough to resolve a prerequisite. Values use the documented medium-difficulty likelihood assumptions at 0.9 reliability; they are illustrative, not population-calibrated. The trusted prerequisite is Force and Pressure → force as a push or pull. Figure 7 Interactive illustration of the two-probe resolution rule on a synthetic Friction scenario. Demonstration only; not a result and not a student record. Randomly splitting individual questions or edges would allow near-duplicate chapter structure to leak across partitions. Both pipelines therefore group examples by target chapter, with zero overlap between train, validation, and test partitions. The policy test contains 11,237 candidates in 930 decision groups from seven held-out chapters. Those groups reuse 31 state templates, so results are reported as benchmark coverage rather than 930 independent learner trials. Figure 8 Chapter-disjoint split proportions for both pipelines. No target chapter appears in more than one partition. Table 4 Chapter-disjoint split sizes. The split manifests record empty pairwise target-chapter overlap for train/calibration/test. Within the policy test, 581 of 930 groups have multiple reference-best candidates. The deployed surrogate selected a reference-optimal candidate in all 930 held-out groups, with mean regret 0 and no infeasible selections. Deterministic baselines also perform strongly, including 930/930 agreement for the explicit reference formula and equal weights. The result establishes reliable execution of the constrained policy; it also shows that the current benchmark does not yet require neural complexity. Verified: the serving artifact, feature ordering, masking, and ranking path are consistent with the reference policy definition on seven held-out chapters. Not measured here: learner answer prediction, causal learning improvement, or optimality of the selected cost weights. Benchmark structure: 930 groups reuse 31 state templates across seven chapters. The chapter-disjoint split tests transfer to unseen curriculum contexts; it is not a study of 930 independent learners. Table 5 Selector comparison on the fixed 930-group test split. Ties count as correct when the selected candidate belongs to the reference-optimal set. Strong simple baselines motivate a deterministic-first implementation. These bars reproduce the recorded engineering acceptance criteria and achieved values. They document the promoted configurations; the selector comparison above provides the appropriate scientific context. Figure 9 Recorded operational criteria versus achieved values for the prerequisite and next-question components. A useful reference method separates verified engineering properties from prospective research claims. The next stage is to connect this reproducible policy to expert labels, calibrated learner evidence, and measured learning outcomes. Established today: 523 prerequisite links survived the operational evidence rules. Next validation: compare the graph with a blinded, multi-expert prerequisite benchmark to measure agreement beyond the current weak labels. Established today: the serving selector reproduced the reference decision rule across all 930 held-out groups and never crossed the feasibility mask. Next validation: introduce harder counterfactual candidates that distinguish the weighted rule from simpler selectors. Established today: every answer produces a deterministic, inspectable state update. Next validation: fit the likelihood parameters and resolution threshold on representative learner response data. Established today: the method is technically reproducible without student data. Next validation: a prospective learner study comparing diagnostic efficiency, gap identification, and subsequent learning outcomes against a fixed-question baseline. Constrained item selection predates this system [1], as do learned knowledge-state models [2], Bayesian skill models for adaptive testing [3], and semantically grounded exercise recommendation [4]. Calibration metrics also require careful interpretation, especially with small samples and bin-dependent estimators [5, 6]. SwaVid’s contribution is the integrated decision contract: a reviewed NCERT graph defines what may be tested, current-session evidence defines what remains unresolved, a hard mask defines what is valid, and an interpretable cost defines what should be asked next. Candidate paths and questions are tied to reviewed NCERT sources and exact concept evidence, not open-ended generation. Grounded support, trusted questions, graph rules, and release criteria remain inspectable independently of the ranking model. The learner state is updated from answers in the active diagnostic rather than inferred from a hidden historical student profile. Inputs, masks, costs, thresholds, splits, and baseline comparisons are explicit enough for another tutor to reproduce or replace each component. Table 6 Primary literature connecting the method to constrained adaptive testing, knowledge-state estimation, exercise recommendation, and probability calibration. Define the eligible conceptual frontier before ranking or generating questions. Fluency cannot repair a concept mismatch. Convert the latest answers into an explicit evidence state. The same question can have different value in different states. Remove untrusted, off-concept, or over-budget questions first. A scoring model should never be able to rescue an invalid action. Begin with the deterministic reference rule. Add a learned residual only when harder data or learner outcomes show a repeatable advantage. The distinctive contribution is the full, inspectable chain: NCERT evidence → reviewed prerequisite graph → current-session evidence update → hard feasibility mask → minimum-cost question selection. Its value is not that every component is unprecedented; its value is that the decision can be reproduced, audited, and improved without surrendering curriculum control to generation. These values make the reported policy reproducible. They are a starting configuration, not universal constants for every tutor or learner population. Table 7 Selected hyperparameter values of the reported configurations, consolidated from the training reports. The policy predicts which feasible question should be asked next—not whether the learner will answer correctly. It ranks trusted questions for the current concept and evidence state, then selects the minimum-cost option. No. The reported policy can operate from answers in the active diagnostic session. A prior can be supplied, but the selection loop does not require names, historical profiles, or stored answer traces. A single response may reflect a slip, guess, or misread. Requiring repeated evidence makes the resolution rule more conservative; the exact threshold should be calibrated for the learner population in any deployment. No. The online selector consumes a published curriculum graph and a trusted question bank. New graph relationships follow a separate offline evidence and publication process. No. A deterministic weighted policy is the reference implementation and remains a strong baseline. A learned residual is useful only when a harder, outcome-linked benchmark demonstrates a measurable advantage. Posterior uncertainty is an explicit cost component, so questions that can resolve the decision remain competitive. If evidence still does not cross a resolution threshold, the system continues probing within its budget rather than pretending certainty. The graph identifies what matters, the learner’s answers identify what remains unresolved, the mask protects validity, and the cost function selects the next question. ## Key Links - [Explore the concept map](https://swavid.com/concept-map) - [Stocking & Swanson (1993), A Method for Severely Constrained Item Selection in Adaptive Testing](https://journals.sagepub.com/doi/10.1177/014662169301700308) - [Piech et al. (2015), Deep Knowledge Tracing](https://papers.nips.cc/paper/5654-deep-knowledge-tracing.pdf) - [Plajner & Vomlel (2016), Student Skill Models in Adaptive Testing](https://proceedings.mlr.press/v52/plajner16.html) - [Ozyurt et al. (2025), Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing](https://papers.nips.cc/paper_files/paper/2025/hash/13707aad517ddd6c09ea02e0f55e1e7a-Abstract-Conference.html) - [Zhang, Kailkhura & Han (2020), Mix-n-Match: Ensemble and Compositional Methods for Uncertainty Calibration](https://proceedings.mlr.press/v119/zhang20k.html) - [Arrieta-Ibarra et al. (2022), Metrics of Calibration for Probabilistic Predictions](https://www.jmlr.org/papers/v23/22-0658.html) - [Open the concept universe](https://swavid.com/concept-map) - [Read Diagnosis before generation](https://swavid.com/research/why-diagnosis-first)