EduResearchBench: A Hierarchical Atomic Task Decomposition Benchmark for Full-Lifecycle Educational Research

TL;DR

This paper introduces EduResearchBench, a comprehensive benchmark for evaluating educational scholarly writing using the Hierarchical Atomic Task Decomposition (HATD) framework. HATD splits the full research lifecycle into six modules and 24 atomic tasks, enabling fine-grained diagnostics beyond holistic scoring. The authors build an automated evaluation pipeline with expert-crafted, per-task prompts, dual-model critique/judging, and a High-Standard data-sourcing process to train EduWrite via supervised fine-tuning on 11K high-quality instruction pairs derived from 55K raw samples. Experimental results show that EduWrite outperforms larger open-source models and approaches frontier closed-source systems in several modules, with curriculum learning further boosting performance especially in policy-oriented tasks. The work demonstrates the importance of data quality density and staged curriculum design for vertical-domain writing tasks and provides a scalable framework for diagnosing and improving complex scholarly writing in education.

Abstract

While Large Language Models (LLMs) are reshaping the paradigm of AI for Social Science (AI4SS), rigorously evaluating their capabilities in scholarly writing remains a major challenge. Existing benchmarks largely emphasize single-shot, monolithic generation and thus lack the fine-grained assessments required to reflect complex academic research workflows. To fill this gap, we introduce EduResearchBench, the first comprehensive evaluation platform dedicated to educational academic writing. EduResearchBench is built upon our Hierarchical Atomic Task Decomposition (HATD) framework, which decomposes an end-to-end research workflow into six specialized research modules (e.g., Quantitative Analysis, Qualitative Research, and Policy Research) spanning 24 fine-grained atomic tasks. This taxonomy enables an automated evaluation pipeline that mitigates a key limitation of holistic scoring, where aggregate scores often obscure specific capability bottlenecks, and instead provides fine-grained, diagnostic feedback on concrete deficiencies. Moreover, recognizing the high cognitive load inherent in scholarly writing, we propose a curriculum learning strategy that progressively builds competence from foundational skills to complex methodological reasoning and argumentation. Leveraging 55K raw academic samples, we curate 11K high-quality instruction pairs to train EduWrite, a specialized educational scholarly writing model. Experiments show that EduWrite (30B) substantially outperforms larger general-purpose models (72B) on multiple core metrics, demonstrating that in vertical domains, data quality density and hierarchically staged training curricula are more decisive than parameter scale.

Figures (4)

• Figure 1: Overview of the EduResearchBench framework. GPT-5 and task templates are used to generate Q-A pairs from processed literature and public data. High-quality samples (score $\geq$ 4) are used to fine-tune Qwen3-30B-A3B, while a randomly selected subset serves as the test set for performance evaluation.
• Figure 2: The HATD Taxonomy of EduResearchBench. It consists of 6 specialized research modules and 24 fine-grained atomic tasks.
• Figure 3: Distribution of Token Lengths in Training Set. The histogram shows the distribution of token lengths (on a logarithmic scale) with the minimum token length of 472, the average of 3,136, and the maximum of 108,520.
• Figure 4: Overall performance of different models. The models are evaluated based on their overall scores, showing the comparison between closed-source models, open-source models, and capability-enhanced LLMs.