Most AI work for education focuses on getting the right answer; classroom teaching requires a different skill set. OmniEdu reframes supervision around four educational capabilities so a single model can not only solve K–12 problems but also locate those problems in a curriculum, infer student misconceptions, and choose appropriate instructional actions.
Key Findings
-
Capability-oriented data mix: The authors construct a 69,999-example instruction-tuning mixture (15.96M supervised response tokens) drawn from 100+ educational and general instruction sources, explicitly labeling examples to teach subject competence, curriculum grounding, diagnostic reasoning, and pedagogical scaffolding — so models learn behaviors, not only answers.
-
Effective at multiple scales: Fine-tuning at 4B/9B/27B shows consistent gains across curriculum-grounding, problem-solving, and tutoring benchmarks; the 27B model achieves notably higher EM/F1 on K12-Bench and strong results on MathFish, EDUMATH, MathTutorBench, and LongTutor — so specialization yields measurable teaching abilities without only increasing model size.
-
Reproducible, audited pipeline: The data pipeline combines deterministic cleaning, LLM-assisted semantic auditing/rewriting, task-specific scoring, token-budgeted diversity selection, and pedagogical instruction assignment—so provenance and pedagogical intent are retained for each example.
Who it's for and tradeoffs
Great fit if you need open, education-focused LLMs that go beyond answer correctness — for researchers building tutoring systems, curriculum-aware assessment tools, or instructional agents that must diagnose and scaffold. Look elsewhere if your priority is raw SOTA on large-scale non-educational benchmarks or if you require multilingual K–12 coverage beyond the released corpus; the work centers on curated K–12 supervision and requires additional localization or expansion for other education systems.
Where it fits
OmniEdu sits between general instruction-tuned LLMs and task-specific tutors: it demonstrates that capability-balanced supervision can adapt general models into systems that reason about curriculum structure and student states while remaining competitive on standard problem-solving benchmarks.
