Key Journals & Conferences

Overview

The publication landscape for LLM-in-education research reveals a fundamental methodological schism in how we study educational technology. On one side, ML/NLP venues (NeurIPS, ACL, TACL) privilege technical contributions—novel architectures, benchmark improvements, and scaling analyses. On the other, educational research journals (Computers & Education, IJAIED) demand pedagogical grounding, learning theory integration, and evidence of educational efficacy. This bifurcation creates a troubling gap: the researchers building LLMs rarely publish in education venues, while those studying educational impact rarely have access to the models they study.

This matters because where research gets published shapes what questions get asked. NeurIPS papers optimize for MMLU scores and benchmark performance—metrics that correlate poorly with tutoring effectiveness. Educational journals emphasize controlled studies with students—but these studies necessarily lag behind rapidly-evolving model capabilities. The result is a literature where technical claims outpace pedagogical validation, and educational critiques address models already superseded. Shahzad et al. (2025) drew from 138 references across this fragmented landscape, illustrating both the interdisciplinary richness and the interpretive challenges of synthesizing work from venues with different epistemological standards.

A critical reading of publication patterns also reveals power asymmetries. High-impact journals like Nature (IF: 64.8) and Science (IF: 56.9) publish LLM breakthroughs almost exclusively from well-resourced labs—OpenAI, Google DeepMind, Anthropic. Educational researchers, typically working with smaller budgets and limited compute, publish in specialized venues with lower visibility. This creates an information asymmetry: claims about LLM educational potential circulate widely in high-prestige venues, while empirical studies of actual educational outcomes remain siloed in specialized journals. Understanding these dynamics is essential for critically evaluating the literature.

Critical Analysis: The Politics of Publication

The Speed-Rigor Tradeoff

The rapid pace of LLM development has created a publication speed crisis that affects both ML and education venues. Conference-driven ML research operates on 4-6 month cycles (submission to publication), while rigorous educational studies—requiring IRB approval, student recruitment, semester-long interventions, and statistical analysis—take 2-3 years. This temporal mismatch means that by the time educational researchers publish studies of GPT-3's classroom impact, the field has moved to GPT-4 and beyond. The result is a troubling pattern: ML venues set the research agenda based on what models can technically do, while educational research perpetually chases a moving target.

Compounding this, the prestige hierarchy of venues creates perverse incentives. Publishing in Nature (IF: 64.8) or NeurIPS counts far more for academic careers than publishing in IJAIED (IF: 4.7) or AIED proceedings. Young researchers face pressure to pursue technically impressive work over pedagogically meaningful work. This isn't merely an academic concern—it shapes which questions get funding, which problems get solved, and ultimately, which students benefit.

Open Access and Knowledge Equity

The access patterns of key venues raise equity concerns. High-impact AI journals like Nature Machine Intelligence and IEEE TNNLS require institutional subscriptions costing thousands of dollars annually—effectively limiting access to well-funded research universities. Paradoxically, the foundational ML work that shapes educational AI remains inaccessible to many educators and education researchers at under-resourced institutions. Open-access venues like JMLR and arXiv provide some counterweight, but peer-reviewed educational technology journals often remain paywalled.

The emergence of open-access education-AI venues like Computers and Education: Artificial Intelligence (launched 2020) represents a deliberate effort to bridge this gap. However, the two-tier system persists: breakthrough model papers appear in high-prestige subscription journals, while applied educational studies circulate in lower-visibility open venues. Critical readers should recognize that the literature they can access may systematically differ from the literature driving field direction.

Methodological Incommensurability

Perhaps most importantly, different venue traditions apply fundamentally different standards of evidence. ML venues evaluate papers primarily on benchmark performance—reproducible, quantitative, and comparable. Educational venues prize ecological validity—does the intervention work in real classrooms with real students? These standards are not merely different; they can conflict. A tutoring system that achieves state-of-the-art benchmark scores may fail when deployed with actual students who get frustrated, distracted, or confused in ways benchmarks don't capture.

This incommensurability explains why claims in the educational AI literature can seem contradictory. A NeurIPS paper might report that an LLM "achieves expert-level performance on mathematics tutoring," while an AIED study finds the same system "fails to support metacognitive development." Both may be correct—they're measuring different things. Sophisticated readers must triangulate across venues, recognizing that technical capability claims (ML venues) and educational efficacy claims (education venues) require different kinds of evidence and should be evaluated by different standards.