Large language models for teaching and learning in higher education: opportunities, challenges, and future directions

Synthesis: Li and colleagues (2026) introduce a special issue of five contributions on large language models in higher education, and the collection's unifying argument is that educational value cannot be inferred from generative capability. Each paper situates an LLM inside an existing educational system: course recommendation, real-time prompting support, automated essay scoring, verifiable online education via blockchain provenance, and knowledge-augmented form filling for institutional services. The editorial's own framing is that LLMs are becoming components of educational systems rather than standalone tools, and its repeated caution is evidentiary — most of the contributions report feasibility or Benchmark performance rather than evidence of learning. The gap the editors identify is between technical performance and educational effectiveness: the questions that matter most for practice, concerning learner development, authentic settings and institutional responsibility, remain insufficiently examined.

Key Findings

1. The collection spans five points in the educational system. Course recommendation (LLM-CR), real-time prompting coaching (ARPG+), essay scoring (RACES), blockchain-verified online education, and knowledge-augmented form filling (KAFA) are presented as components rather than as competing approaches, which is the editorial's central organizing claim.
2. Selective integration outperforms wholesale replacement. LLM-CR runs its expensive LLM processing offline and reuses the resulting semantic summaries, producing modest online overhead. The editors use this as an illustration of integrating LLM capability only where it adds value to an existing pipeline.
3. Coaching architectures aim at their own withdrawal. ARPG+ uses a dual-path design — lightweight analysis for routine interactions, deeper analysis at diagnostically important moments — and reports a higher simulated no-hint rate in the final five turns than fixed-support strategies. The design goal is fading assistance, not perpetual support.
4. Scoring gains are reported without significance testing. RACES reports higher Quadratic Weighted Kappa and higher SimCSE-based similarity values than pretrained and fine-tuned baselines on ASAP 2.0, but the paper reports no multi-seed uncertainty analysis or statistical-significance tests. The editors do not upgrade that claim on the paper's behalf.
5. Auxiliary Feedback is proxy-generated. The feedback comments used in RACES's reward modeling were generated by Qwen2.5-Instruct-32B and treated as proxy feedback, while ASAP 2.0 scores were the only human-assigned labels. The paper positions the system as an auxiliary tool for preliminary scoring and workload reduction, not as an autonomous replacement for human raters.
6. Provenance is treated as a first-class requirement. The consortium-blockchain contribution proposes tamper-evident records for educational events, AI outputs, grading actions and certificates on a four-peer Hyperledger Fabric network, so that institutions can reconstruct how consequential decisions were made and who held authority to confirm or revise them. Its prototype grader used keyword matching and its interactions were scripted, so it provides feasibility evidence only.
7. Simulation and proxy-domain evidence are the collection's evidentiary ceiling. ARPG+ was evaluated with LLM-based simulated learners, and KAFA on 1,301 form fields across the public recruitment sites of 44 companies as a proxy for admissions and registration forms. Both papers are explicit that these results do not establish classroom effectiveness or performance on authentic higher-education platforms.

The gap between benchmark performance and educational value

The editorial's most useful contribution is its insistence on separating what a system can do from what it does educationally. LLM-CR improves next-course ranking on six XuetangX subsets, but the editors note that better ranking does not establish better learning outcomes or curriculum quality. RACES improves agreement with reference scores on a benchmark, but the reference scores are the only human labels in the design and the auxiliary feedback is model-generated. KAFA improves field-filling accuracy on recruitment websites, but recruitment forms are not university admissions systems. In each case the reported result is real and the inferential distance to educational value is also real. The editors are explicit that the collection's contributions "illustrate" rather than "establish," and that the shared implication is that value depends on how LLMs are situated within relationships among people, knowledge and institutional practice.

Scaffolding, cognitive load, and the fading problem

ARPG+ is the collection's most theoretically developed contribution, and it is also the one whose claims are most carefully bounded. It integrates cognitive load theory, the zone of proximal development, adaptive scaffolding and metacognitive support into a real-time coaching framework for prompting — treating prompting itself as the skill to be developed rather than as a means to an end. The dual-path architecture is a design response to the cost of continuous deep analysis, and the reported no-hint rate in the final five turns is a direct operationalization of fading. Because the learners were simulated, the result speaks to whether the architecture behaves as intended, not to whether students learned to prompt better. That distinction matters for anyone tempted to read the simulated prompt-quality gains as classroom effects, and the editors' framing of the collection as a whole depends on readers keeping it.

Institutional accountability when AI enters assessment

The blockchain contribution addresses a problem the other papers leave open: once an LLM contributes to grading or certification, institutions need a way to reconstruct the decision path. The editorial treats this as an accountability requirement rather than a technical flourish, noting that when AI becomes embedded in Assessment and certification, mechanisms are needed not only for improving outputs but for establishing who held authority to confirm or revise them. The prototype is weak as evidence — scripted interactions, keyword-matching grader, no authentic institutional AI Governance — but the framing is the contribution. It also connects directly to the collection's recurring concern about trust and to the privacy questions raised by knowledge-augmented agents that maintain personal knowledge bases, which KAFA's authors identify as an unresolved deployment requirement alongside institutional terminology, access control and human confirmation.

What this means for practice

• Instructors. The collection's most transferable lesson is architectural: integrate LLM capability at specific points in an existing workflow rather than replacing the workflow. LLM-CR's offline-processing design is the template.
• Assessment designers. Treat automated scoring as auxiliary until repeated-scoring consistency, paraphrase robustness and subgroup fairness have been tested. RACES's own authors identify those as open priorities, and the benchmark result does not cover them.
• Educational technology developers. Design for fading. ARPG+'s explicit goal of withdrawing hints is a more demanding specification than maximizing support, and it is the one that aligns with developing learner capability.
• Institutional leaders. Provenance and human-approval mechanisms are governance requirements, not features. If AI touches grading or certification, plan how a decision can be reconstructed and who is accountable for confirming it.
• Researchers. Report model versions, prompts, retrieval sources and interaction histories. The editorial's critique of the field is that these are routinely under-specified, which prevents cumulative evidence.

Limitations

• Most contributions report feasibility, not effectiveness. ARPG+ used simulated learners and KAFA used recruitment websites as a proxy domain, so neither establishes classroom or institutional impact.
• Statistical rigor varies across the collection. RACES reports no multi-seed uncertainty analysis or significance tests, and the editors present its scoring gains as benchmark agreement rather than as validated reliability.
• The special issue is small and thematically bounded. Five contributions cannot represent the breadth of LLM use in higher education, and all five concern system design rather than learner outcomes.
• The blockchain contribution is a scripted prototype. With a keyword-matching grader and no authentic institutional governance, it demonstrates feasibility only.
• The editorial's own claims are programmatic. Its argument that future research should establish when, for whom and under what conditions LLM use produces educational value is a research agenda, not a finding from the collection.
• Equity and accessibility are named as priorities rather than examined. The future-directions section identifies them as needed work, which means the collection does not yet report evidence on distributional effects.

Citation

Li, Z., Conesa, J., Gabbouj, M., et al. (2026). Large language models for teaching and learning in higher education: opportunities, challenges, and future directions. International Journal of Educational Technology in Higher Education, 23(1).