The Order Matters: Sequential Fine-Tuning of LLaMA for Coherent Automated Essay Scoring

Technical Summary: The Order Matters: Sequential Fine-Tuning of LLaMA for Coherent Automated Essay Scoring

Problem Statement
Automated Essay Scoring (AES) systems face a critical challenge in modeling the interdependent nature of discourse elements within an essay. Traditional approaches often treat components such as the lead, position, claim, evidence, and conclusion in isolation. This fragmentation fails to capture the logical flow and hierarchical dependencies inherent in argumentative writing (e.g., how the clarity of a position influences the strength of a claim). Consequently, existing models struggle with coherence, generalization across diverse writing styles, and the reliable assessment of integrative components. Furthermore, while Large Language Models (LLMs) offer potential, there is a lack of understanding regarding whether smaller, efficiently fine-tuned models can outperform massive, general-purpose baselines when task dependencies are explicitly modeled.

Methodology
The study investigates the impact of training curricula on the fine-tuning of LLaMA-3.1-8B for AES. The authors utilize the PERSUADE 2.0 corpus, a dataset of over 25,000 argumentative essays annotated with five specific discourse components: Lead, Position, Claim, Evidence, and Conclusion.

To adapt the model efficiently, the authors employ Low-Rank Adaptation (LoRA) with 4-bit quantization (NF4), allowing training on a single A100 GPU while preserving pre-trained knowledge. The core experimental design compares three distinct training curricula against a zero-shot LLaMA-70B baseline:

1. Independent Fine-Tuning: Five separate models are trained, each dedicated to a single discourse component using task-specific LoRA adapters. This serves as a baseline to assess the value of shared representations.
2. Randomized (Multi-Task) Fine-Tuning: A single model is trained on a shuffled mixture of all five component datasets simultaneously. This approach treats tasks as independent and identically distributed, ignoring sequential structure.
3. Sequential (Curriculum) Fine-Tuning: The primary proposed method. A single model is fine-tuned progressively following the natural logical order of essay writing: Lead $\rightarrow$ Position $\rightarrow$ Claim $\rightarrow$ Evidence $\rightarrow$ Conclusion. Parameters learned from one stage serve as the initialization for the next, aiming to build complex representations by leveraging knowledge from foundational components.

Key Results
Experiments on the PERSUADE 2.0 test set reveal that the structure of the training curriculum significantly impacts model performance:

• Superiority of Sequential Learning: The Sequential curriculum yielded the strongest overall results. It achieved the highest F1 scores on three of the five components: Claim (40%), Evidence (65%), and Conclusion (87%). Notably, it achieved an accuracy of 85% on the Conclusion task.
• Task Interdependence: The Sequential model significantly outperformed the Independent models, particularly on the Conclusion task (87% F1 vs. 12% F1 for Independent). This suggests that evaluating integrative components requires contextual representations informed by prior essay sections, which isolated training fails to capture.
• Inconsistency of Randomized Learning: While the Randomized approach improved Position scoring (57% F1), it performed poorly on the Lead component (7% F1), indicating that unstructured multi-task learning is less robust than a curriculum aligned with discourse structure.
• Small vs. Large Models: A compact, task-optimized LLaMA-8B model (Sequential) outperformed the much larger, zero-shot LLaMA-70B baseline on context-dependent tasks. Specifically, the 8B model surpassed the 70B baseline on Evidence (65% vs. 49% F1) and Conclusion (87% vs. 60% F1), despite the 70B model's superior capacity.

Significance and Claims
The paper claims that the "order matters" in fine-tuning, asserting that curriculum design aligned with the intrinsic structure of the task acts as a powerful inductive bias. The authors argue that:

1. Curriculum Design is Critical: Modeling task dependencies through sequential fine-tuning materially improves AES performance, particularly for integrative components where logical flow is essential.
2. Efficiency and Competitiveness: Small, task-optimized models can be competitive with, and even surpass, substantially larger LLMs when the training strategy is aligned with the task's hierarchical nature. This offers a practical, cost-effective path for scalable educational assessment without relying on resource-intensive proprietary APIs.
3. Generalizability: By moving away from isolated component scoring, the proposed approach addresses the issue of incoherent feedback, providing a more robust framework for evaluating student writing that reflects the logical relationships between essay sections.

The authors conclude that for structured prediction tasks in NLP, how a model is taught (the curriculum) is as important as what it is taught, enabling the induction of more robust and efficient representations.