PACE: A Personalized Adaptive Curriculum Engine for 9-1-1 Call-taker Training

9-1-1 call-taking training requires mastery of over a thousand interdependent skills, covering diverse incident types and protocol-specific nuances. A nationwide labor shortage is already straining training capacity, but effective instruction still demands that trainers tailor objectives to each trainee’s evolving competencies. This personalization burden is one that current practice cannot scale. Partnering with Metro Nashville Department of Emergency Communications (MNDEC), we propose PACE (Personalized Adaptive Curriculum Engine), a co-pilot system that augments trainer decision-making by (1) maintaining probabilistic beliefs over trainee skill states, (2) modeling individual learning and forgetting dynamics, and (3) recommending training scenarios that balance acquisition of new competencies with retention of existing ones. PACE propagates evidence over a structured skill graph to accelerate diagnostic coverage and applies contextual bandits to select scenarios that target gaps the trainee is prepared to address. Empirical results show that PACE achieves 19.50% faster time-to-competence and 10.95% higher terminal mastery compared to state-of-the-art frameworks. Co-pilot studies with practicing training officers further demonstrate a 95.45% alignment rate between PACE’s and experts’ pedagogical judgments on real-world cases. Under estimation, PACE cuts turnaround time to merely 34 seconds from 11.58 minutes, up to 95.08% reduction.

💡 Research Summary

The paper introduces PACE (Personalized Adaptive Curriculum Engine), a novel co‑pilot system designed to personalize and accelerate the training of 9‑1‑1 call‑takers, a domain that demands mastery of over a thousand interdependent skills across dozens of incident types. The authors first formalize the training problem as curriculum optimization over a structured knowledge graph. Nodes represent fine‑grained competencies (conditions, questions, instructions) and edges encode procedural order, prerequisite, and entailment relationships, yielding a directed graph with 1,053 nodes and 1,283 edges.

PACE maintains a probabilistic belief state over each skill’s mastery using per‑node Beta distributions. After every simulated call, an LLM parses the transcript and extracts structured observations (compliance, partial compliance, violation, or not applicable) together with error types (slip, misconception, omission). These observations update the Beta parameters with weighted pseudo‑counts that reflect the severity of the error. To address the sparsity of direct observations, the system pre‑computes a similarity index between skills based on sentence‑transformer embeddings and graph depth, allowing evidence from a observed skill to be propagated to similar, unobserved skills.

Beyond static belief tracking, PACE models individual learning dynamics. The learning pace λ is defined as the average marginal mastery gain per practice opportunity, while forgetting is captured by a power‑law decay model θ(τ+Δτ)=θ(τ)·(1+κ·Δτ)^{‑ψ}, where ψ is a learner‑specific forgetting rate. Both λ and ψ are estimated from the trainee’s interaction history during an initial cold‑start phase of 15 trainer‑assigned scenarios.

Scenario selection is cast as a contextual bandit problem. The context vector at session t aggregates (i) current posterior means and variances for all skills, (ii) the learner’s λ and ψ, and (iii) the similarity matrix. Using a Thompson‑sampling‑style policy, PACE samples a batch of K scenarios that jointly maximize expected learning gain (acquisition of new skills) and retention (mitigating forgetting), while also exploring uncertain skills.

Empirical evaluation combines offline simulation against state‑of‑the‑art knowledge‑tracing baselines (DKT, AKT, etc.) and a real‑world co‑pilot study with the Metro Nashville Department of Emergency Communications. Results show that PACE reduces time‑to‑competence by 19.5 % and improves final mastery scores by 10.95 % relative to the baselines. In the field study, PACE’s scenario recommendations matched expert trainers’ choices 95.45 % of the time, confirming that the learned policy captures domain‑relevant pedagogical heuristics. Moreover, the system cuts the average simulation‑to‑debrief turnaround from 11.58 minutes to just 34 seconds—a 95.08 % reduction—potentially saving thousands of trainer hours per training cohort.

The authors discuss broader implications: PACE demonstrates how a combination of graph‑structured knowledge, Bayesian belief updating, individualized learning/forgetting dynamics, and contextual bandits can address high‑stakes, large‑scale personalized education problems. Limitations include reliance on text‑based transcripts, independent per‑skill belief updates, and the need for more interpretable policy explanations. Future work is suggested on multimodal feedback integration, richer probabilistic graphical models for skill interdependencies, and scalable dashboards for trainer oversight. Overall, the paper provides a compelling, data‑driven solution that could be adapted to other complex domains such as medical education, aviation, or military training.