This preprint formalizes Reconstruction Lower Bounds for Learner Models (RLB-LM), identifying fundamental, identifiability-driven limits on what can be inferred about latent learner or conceptual models from behavioral observations. The work establishes non-asymptotic impossibility results that apply across broad classes of probing strategies and evaluation systems. RLB-LM provides a theoretical foundation for understanding the limits of assessment, evaluation, and probing in AI and educational settings and serves as a reference point for system design under unavoidable uncertainty. This record establishes authorship and priority for the RLB-LM framework.
Murad Ahmadov (Fri,) studied this question.