Long-term tutoring stops guessing from retrieved history: evidence, interpretation and teaching policy become one inspectable chain, lifting evidence accuracy by 13.2 points on average

Long-term personalized tutoring now has an explicit chain from stored history to instructional action, with teaching policies that update as new observations arrive and keep links to the evidence that supports them. Earlier LLM tutoring systems used memory, retrieval and learner profiles for personalization, but retrieving relevant records does not by itself specify how those records should guide teaching. On LongTutor's 1,000-case human-annotated split, nine fixed-backbone comparisons raise evidence accuracy by 13.2 percentage points and diagnosis accuracy by 11.6 points on average, with six configurations showing evidence gains whose intervals exclude zero.

The system first resolves history into evidence with source identifiers, then derives a learner interpretation and teaching policy, and finally compiles them into learner-facing instructional directives. Evidence resolution runs through the Evidence Resolution Loop, where questions spanning several sessions use mapper calls to select observations from history chunks and resolver calls combine them into answers with source identifiers. Interpretation follows the CAMM framework from self-regulated learning research, representing cognitive, affective, metacognitive and motivational processes separately, and each interpretation records its supporting observations and inference boundaries.

An offline stage, PairRoute, uses intermediate execution records to separate interface failures in evidence access from model execution failures, then builds stage-level supervision for confirmed model failures. An incorrect final response alone does not reveal whether the failure comes from the interface's access to or representation of evidence, or from the model failing to execute a feasible stage. On a 100-case development set, evidence accuracy under the selected interface rises from 212/300 to 250/300 across three model-training rounds, with the first update gaining 6.0 percentage points and a paired interval of [1.7, 10.3].

The framework is evaluated on LongTutor's 1,000-case human-annotated split, where each case carries 199 history records and a separate current question, covering evidence acquisition, learner diagnosis and teaching; interface revision and model training use a 369-case discovery set, a 32-case probe set and a 100-case development set. It targets researchers and engineers building tutoring systems that must turn long learning histories into instructional decisions, and the evaluation unit is an independent history snapshot.

LongTutor evaluates independent history snapshots and does not measure real learners' retention, transfer or progress over time; teaching scores rely on model judges, with some configurations sharing the evaluator's model family. The refinement study uses one local backbone, three supervised rounds and an adaptively reused development set, and the crossed model-interface evaluations characterize compatibility without budget-matched controls for alternative refinement strategies. The interface represents learning-related evidence through engineering heuristics; it does not provide calibrated estimates of psychological traits or establish causal effects of instructional interventions.