DeepEdu measures tutoring-system engineering, not student learning gains
The reported retrieval and agentic improvements belong to specific system benchmarks; educational and legal conclusions need separate evidence.
DeepEdu-v1 targets Vietnamese educational content with a local tutoring system. Its SCALE framework combines a long-context inference design with an agentic playbook built from verified interaction evidence rather than repeated model fine-tuning.
The authors report 7.7 times fewer retrieval calls, roughly 35% lower prefill latency, and about twofold TTFT improvement over standard vLLM in the deployed configuration. Agentic benchmark accuracy changes from 70.0% to 79.5%. These are not controlled measurements of student learning.
The paper motivates self-hosting partly through an asserted data-sovereignty concern. A research paper’s description of a law is not an independent legal determination. A school considering the system needs its own data-handling assessment alongside engineering tests.
The practical comparison is the same curriculum workload under the existing serving stack and the proposed inference/playbook design. Measure response quality, TTFT and the process that accepts new playbook entries before claiming educational benefit.
Editorial correction (2026-10-13 UTC): Separated engineering benchmarks from educational outcomes and attributed legal motivation.