Beyond Score Prediction: LLM-Based Essay Scoring and Feedback Generation via Reinforcement Learning with Rubric Rewards

2026-07-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, quick

Summary

RLAES is a unified large language model (LLM) framework designed for automated essay scoring (AES) and feedback generation (AFG), leveraging reinforcement learning (RL) post-training. Addressing limitations in existing prompt engineering or supervised fine-tuning approaches, RLAES introduces Rubric-based Feedback Evaluation (RFE), an essay-grounded framework with 166 fine-grained binary rubric items and an LLM-as-judge, to measure and interpret feedback quality. Building on RFE, Adaptive Gated Feedback Optimization (AGFO) activates rubric-based feedback rewards on demand during RL, reducing evaluation overhead and improving feedback quality. Additionally, Adjacent Contrastive Reasoning (ACR) enhances ordinal score calibration by contrasting adjacent score levels. Experimental results on the ASAP benchmark show RLAES-AGFO achieves a QWK of 0.803, outperforming other LLM-based methods in scoring, while maintaining feedback quality comparable to GPT-5.5 and preventing feedback degradation seen in score-only RL. Code and datasets are publicly available.

Key takeaway

For NLP Engineers developing automated essay scoring or feedback systems, RLAES offers a robust RL-based approach. You should consider integrating rubric-based feedback evaluation (RFE) and adaptive gated optimization (AGFO) to achieve high scoring accuracy (QWK = 0.803) while ensuring feedback quality comparable to advanced models like GPT-5.5. This framework helps avoid feedback degradation often seen in score-only RL systems, providing a unified solution for both tasks.

Key insights

RLAES unifies LLM-based essay scoring and feedback generation using reinforcement learning with rubric-based rewards for measurable quality.

Principles

• Rubric-based evaluation makes feedback measurable.
• On-demand reward activation reduces RL overhead.
• Explicitly contrasting scores improves calibration.

Method

RLAES jointly optimizes scoring and feedback via RL, using RFE for feedback evaluation, AGFO for adaptive rubric rewards, and ACR for ordinal score calibration.

In practice

• Apply RFE for interpretable feedback quality.
• Use AGFO to optimize RL feedback rewards.
• Implement ACR for better score calibration.

Topics

• Large Language Models
• Automated Essay Scoring
• Reinforcement Learning
• Feedback Generation
• Rubric-based Evaluation
• ASAP Benchmark

Code references

• hellomuyi/RLAES

Best for: AI Engineer, Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Beyond Score Prediction: LLM-Based Essay Scoring and Feedback Generation via Reinforcement Learning with Rubric Rewards — The RLAES framework unifies LLM-based essay scoring and feedback generation via reinforcement learning with rubric-based rewards, outperforming prior methods.
• Rubric-Based Rewards for RL — Rubric-based rewards extend RL to subjective LLM tasks by providing structured, interpretable evaluation criteria.
• Feedback-to-Rubrics: Can We Learn Expert Criteria from Inline Comments? — Inline comments can be distilled into reusable natural-language rubrics, enhancing LLM-driven writing and review support.
• WrAFT: a Modularized Automated Writing Evaluation System for Argumentative Essays — WrAFT's modular LLM-based design delivers high accuracy in essay scoring and human-approved comprehensive feedback.
• Building AdvancedIF: Evolving Instruction Following Beyond IFEval and “Avoid the Letter C” — Current instruction-following benchmarks often prioritize programmatic verifiability over real-world utility, hindering AI development.
• RETUYT-INCO at BEA 2026 Shared Task 2: Meta-prompting in Rubric-based Scoring for German — Meta-prompting enables LLMs to dynamically create prompts for rubric-based short answer scoring.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.