RACES (Reward-Aligned Consistent Essay Scoring) framework

Research Method · All research methods
Evidence · 6 claims (3 for, 3 mixed) · 4 studies (2 causal, 2 design), q2 · 1 of 4 report an effect size · 6 claims rest on one study

Description

RACES (Reward-Aligned Consistent Essay Scoring) is a unified AES framework that "integrates LoRA-based parameter-efficient fine-tuning, reward modeling, and PPO with KL regularization". It runs an offline inference–feedback–optimization loop: LoRA SFT establishes basic score-label mapping with standardized prompts; a reward model learns ranking preferences between preferred proxy responses and harder negative samples; PPO with KL divergence and critic constraints optimizes the policy while limiting distribution drift. Alignment is proxy-based, using LLM-generated feedback rather than human-written comments.

Accounts

• RACES: a reward-aligned, consistency-regularized framework for LLM-based automated essay scoring: RACES (Reward-Aligned Consistent Essay Scoring) is a unified AES framework that "integrates LoRA-based parameter-efficient fine-tuning, reward modeling, and PPO with KL regularization". It runs an offline inference–feedback–optimization loop: LoRA SFT establishes basic score-label mapping with standardized prompts; a reward model learns ranking preferences between preferred proxy responses and harder negative samples; PPO with KL divergence and critic constraints optimizes the policy while limiting distribution drift. Alignment is proxy-based, using LLM-generated feedback rather than human-written comments. (Zhenxin Zhang et al. (2026))

Claims

• RACES configurations improve QWK essay-scoring agreement over evaluated pretrained baselines on ASAP 2.0, with LLaMA2-7B-based configurations slightly outperforming Qwen2.5-Instruct-7B-based ones [+M]
• Mistral-7B reward model outperforms GPT2-XL on Hit@k ranking and Pairwise Accuracy for proxy preference discrimination, at higher computational cost [+M]
• Anchored judge prompts restore score discrimination and raise the reward's agreement with independent human ratings (Spearman ρ 0.672→0.741) [+W]
• After PPO optimization, auxiliary feedback semantic similarity (SimCSE) reaches 81.3% and 79.3% for the two policy models, higher than evaluated pretrained baselines [~W]
• Base-model size effects persist after LoRA adaptation: the 7B model gained more on ROUGE-2 and ROUGE-L, yet 3B-r16 outperformed 7B-r4 on ROUGE-L and BERTScore-F1 [~W]
• One LoRA adapter fine-tuned on the two exams' pooled ~3,900 dual-graded examples brings five small open models to parity or better with a human grader and makes them insensitive to harsh personas [~W]

Related Research Methods

-

Key Sources

• Zhenxin Zhang, Ziyu Ding, Mengyun Liu and Haiwei Sang. (2026). RACES: reward-aligned consistent essay scoring with large language models. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00607-8

Page metadata

Field	Value
type	research-method
id	races-reward-aligned-consistent-essay-scoring-framework
title	RACES (Reward-Aligned Consistent Essay Scoring) framework
description	RACES (Reward-Aligned Consistent Essay Scoring) is a unified AES framework that "integrates LoRA-based parameter-efficient fine-tuning, reward modeling, and PPO with KL regularization".
status	review
generated	by process:settle-candidates on 2026-10-12
trust	unverified — no one has confirmed this page's content

Frontmatter is the page's machine-readable record — see the Schema & Guide.