A Unified Framework to Elicit Structured Feedback for Interpretable Multi-Trait Essay Scoring

Shihang Yang; Sanwoo Lee; Ningning Zhao; Yunfang Wu

cs.CLAug 28, 2026

A Unified Framework to Elicit Structured Feedback for Interpretable Multi-Trait Essay Scoring

Authors: Shihang Yang, Sanwoo Lee, Ningning Zhao, Yunfang Wu

Organizations: National Key Laboratory for Multimedia Information Processing, Peking University · School of Computer Science, Peking University · School of Chinese Language and Literature, Beijing Normal University

Abstract

Multi-trait Automated Essay Scoring (AES) requires rubric-grounded reasoning across interdependent traits, rather than isolated score prediction. Existing feedback-enhanced methods often decouple feedback from scoring or assess traits independently, weakening score--feedback consistency and rubric alignment. We propose HiFTS, a unified autoregressive framework that generates hierarchical CoT feedback before predicting trait-level and holistic scores. HiFTS distills rubric-grounded hierarchical CoT feedback from a teacher LLM and trains student models to jointly generate feedback and scores. HiFTS further applies Group Relative Policy Optimization with a composite reward balancing score agreement, calibration, feedback quality, and structural validity. At inference, a lightweight global prior provides holistic guidance to reduce drift during long-form reasoning. We also introduce CFMS-34, a Chinese multi-trait AES dataset with 951 essays annotated with holistic scores and 34 rubric-based traits. Experiments on CFMS-34 and ASAP++ show that HiFTS achieves strong holistic and trait-level scoring while producing coherent, rubric-aligned feedback.

Figures & tables

Dimension	Code	Definition / Trait Description
Content	C01–C03	Authenticity, sufficiency, and novelty of selected materials.
	C04–C05	Alignment with the central theme; depth of thought and originality.
	C06	Multi-perspective thinking and appropriateness of material presentation.
Structure	S01	Logical paragraph division and structural clarity.
	S02–S03	Use of foreshadowing, transitions, and cohesive devices.
	S04–S05	Structural coherence (e.g., “General-Specific-General”) and thread development.

Table 1: Definition of scoring traits in the CFMS-34 dataset.

Figure 1: Distribution of score differences between the two raters.

Statistic	Value
Total Essays	951
Language	Chinese (Simplified)
Raters per Essay	2 (Human Experts)
Avg. Essay Length	492.4 Chinese characters
Score Dimensions	Overall + 34 Sub-traits
Score Range	0.0 – 5.0

Table 2: Statistics of the CFMS-34 dataset.

Dataset	Language	Expert Annotation	Dual Raters	Multi-trait	Fine-grained
EFCAMDAT Geertzen et al. (2013)	EN	$\times$	$\times$	$\times$	$\times$
ICLE Granger et al. (2009)	EN	$\checkmark$	$\times$	$\checkmark$	$\times$
TOEFL11 Blanchard et al. (2013)	EN	$\checkmark$	$\times$	$\times$	$\times$
FCE Yannakoudakis et al. (2011)	EN	$\checkmark$	$\checkmark$	$\times$	$\times$
ASAP++ Mathias and Bhattacharyya (2018)	EN	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
TOREE Zhuang et al. (2024)	CN	$\checkmark$	$\times$	$\checkmark$	$\times$

Table 3: Comparison with existing essay scoring datasets.

Model	Mean Score	QWK
GPT-4.1 OpenAI (2025)	3.3475	0.4222
Gemini-3 Google (2025)	3.1937	0.4363
Doubao-Seed-1.6 ByteDance (2025)	2.0743	0.1595
Qwen3-max Qwen Team (2025)	3.4059	0.4201
DeepSeek-V3 DeepSeek-AI (2024)	3.5792	0.4130
DeepSeek-R1 DeepSeek-AI (2025)	3.0396	0.3271

Table 4: LLM performance on CFMS-34.

Figure 2: Overview of HiFTS. HiFTS first constructs hierarchical CoT feedback with an LLM teacher, then trains the student model with SFT and GRPO. During inference, a BERT-based global prior guides feedback generation and score prediction for more consistent multi-trait AES.

Method	Explain.	CFMS-34				ASAP++
Method	Explain.	Overall	Traits	MSE	WinRate	Prompts	Traits	WinRate
HISK	$\times$	0.405	0.309	1.374	-	0.644	0.611	-
STL-LSTM	$\times$	0.484	0.365	1.454	-	0.684	0.656	-
MTL-BiLSTM	$\times$	0.519	0.396	1.306	-	0.665	0.638	-
ArTS	$\times$	0.264	0.149	1.960	-	0.717	0.695	-
RMTS	$\checkmark$	0.296	0.303	1.431	-	0.720	0.704	-

Table 5: Main scoring and feedback results on CFMS-34 and ASAP++.

Method	Over.	Cont.	PA	Lang.	Nar.	Org.	Conv.	WC	SF	Style	Voice	Mean
HISK	0.718	0.679	0.697	0.605	0.659	0.610	0.527	0.579	0.553	0.609	0.489	0.611
STL-LSTM	0.750	0.707	0.731	0.640	0.699	0.649	0.605	0.621	0.612	0.659	0.544	0.656
MTL-BiLSTM	0.764	0.685	0.701	0.604	0.668	0.615	0.560	0.615	0.598	0.632	0.582	0.638
ArTS	0.751	0.730	0.751	0.698	0.725	0.672	0.668	0.679	0.678	0.721	0.570	0.695
HiFTS-PPO w/o prior	0.751	0.715	0.730	0.662	0.708	0.645	0.640	0.652	0.625	0.690	0.565	0.671
HiFTS w/o prior	0.754	0.722	0.745	0.685	0.712	0.670	0.635	0.668	0.662	0.688	0.597	0.685

Table 6: Trait-level QWK on ASAP++ with Qwen2.5. The best result for each trait is shown in bold.

Flat feedback (DeepSeek-Reasoner)

Hierarchical CoT feedback (Ours)

This essay revolves around the interaction between “me” and the hamster, with genuine emotions, but it is overall quite simple , lacking in content richness and loose in structure. The opening directly introduces the hamster’s appearance and habits, which is concise but lacks appeal . The main body describes playing, feeding, breeding, and changes in the relationship, but the material is organized haphazardly , with a lack of natural transitions between paragraphs—for example, from “I loved watching it eat sunflower seeds” to “it gave birth to nine babies,” resulting in a logical leap . Detail description is insufficient ; for instance, the process of eating sunflower seeds is attempted but not vivid enough . While emotional expression is sincere, the depth of thought is lacking , failing to explore deeper insights. The ending suddenly confesses feelings , appearing somewhat rushed. The language is fluent but plain, with basically no rhetorical devices or varied expressions. Observation follows a certain order but lacks clear prioritization. Overall …

L1 Global Essay Understanding

\Rightarrow

The student essay “Little Hamster” is …. L2 Dimension Analysis Content

\Rightarrow

The essay includes …These materials are …(C02), and they are …theme (C04). The author … in a …order (E12). In particular , the description of …is vivid and expressive (E11). The essay also elaborates …engaging. However , the essay’s …lacks deeper reflection (C05), and the integration between multi-perspective content and material selection …(C06), while handling of transitional materials appears somewhat rigid (S10). Structure

\Rightarrow

In terms of structural organization, the essay begins with …(S11), and the conclusion remains …(S12). The content of …(S06), and … maintain a connection to topic sentences (S07). The preceding and subsequent content is also presented with …(S02). However , transitions between …(S03), resulting in …(S08). Expression

\Rightarrow

…

Table 7: A qualitative case study comparing flat feedback and hierarchical CoT feedback on the same essay.

Figure 3: Trait-level QWK on CFMS-34, where dashed lines denote averages values.

Method	Grounding
HiFTS-SFT	0.60
HiFTS-PPO w/o prior	0.72
HiFTS-PPO	0.74
HiFTS w/o prior	0.74
HiFTS	0.78

Table 8: Rule-based grounding scores of HiFTS-generated feedback on CFMS-34 with Qwen2.5.

Method	Overall	Traits	WinRate
BERT prior	0.489	-	-
Qwen zero-shot	0.235	0.154	-
Qwen sft w/o CoT	0.432	0.193	-
HiFTS-SFT	0.528	0.387	Ref.
HiFTS-PPO w/o prior	0.572	0.417	0.603
HiFTS-PPO	0.647	0.407	0.651

Table 9: Ablation results on CFMS-34 with Qwen2.5. WinRate is computed against SFT feedback.

Figure 4: Effect of $\alpha$ on holistic and average trait QWK on the CFMS-34 development set.

Window size	Overall	Traits
16	0.612	0.422
32	0.641	0.432
64	0.656	0.439
128	0.659	0.443

Table 10: Ablation of the sliding-window size

h

for the holistic QWK reward on CFMS-34 with Qwen2.5.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Trait	Mean	Std	Exact	Within-1	QWK
Overall	3.64	0.74	0.59	0.97	0.57
C01	4.08	0.86	0.40	0.87	0.25
C02 †	3.96	0.84	0.43	0.87	0.32
C03	3.63	0.81	0.45	0.90	0.34
C04 †	3.81	0.79	0.50	0.89	0.33
C05 †	3.23	0.86	0.49	0.93	0.52

Appendix

Table 11: Per-trait score statistics and inter-rater agreement on CFMS-34 (951 essays).

\dagger

denotes the 20-trait subset used in the main experiments.

Automated Essay Scoring

Automated essay scoring uses computational models to assess student writing against human ratings, rubrics, or proficiency criteria.

See all 56 papers

Beyond Score Prediction: LLM-Based Essay Scoring and Feedback Generation via Reinforcement Learning with Rubric Rewards
Jul 21, 2026Xuefeng Jin, Jiashuo Zhang, Teng Cao +1Automated Essay Scoring Rubric-Based RL
Alibaba Cloud Computing, Alibaba Group
PsyScore: A Psychometrically-Aware Framework for Trait-Adaptive Essay Scoring and ZPD-Scaffolded Feedback
Jun 18, 2026Wei Xia, Jin Wu, Haoran Shi +2Educational Assessment Automated Essay Scoring
Department of Educational Psychology, East China Normal University · 2Shanghai Institute of Artificial Intelligence for Education, East China Normal University · School of Computer Science and Technology, East China Normal University
When Rubrics Change: Cross-Rubric Generalization for Critical Thinking Essay Scoring
Jul 15, 2026Nischal Ashok Kumar, Payu Wittawatolarn, Sana Kang +7Educational Assessment Automated Essay Scoring
University of Massachusetts Amherst, United States · University of Minnesota, United States · Brighter Research, United States +3

Jul 21, 2026J/K move · Enter open · S save

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

Xuefeng Jin, Jiashuo Zhang, Teng Cao, Bin Yang

Alibaba Cloud Computing, Alibaba Group

Automated Essay Scoring Rubric-Based RL Rubric-Based Evaluation

Open paper AskPDF arXiv

Large language models (LLMs) have been widely applied to automated essay scoring (AES) and automated feedback generation (AFG). However, existing studies rely primarily on prompt engineering or supervised fine-tuning, while systematic research on reinforcement learning (RL) post-training and automated evaluation of feedback quality remains limited. We propose RLAES, a unified LLM framework that jointly optimizes essay scoring and feedback generation through RL. To make feedback quality measurable, interpretable, and usable for training, we introduce Rubric-based Feedback Evaluation (RFE), an essay-grounded feedback evaluation framework comprising 166 fine-grained binary rubric items and an LLM-as-judge. Building on RFE, we propose Adaptive Gated Feedback Optimization (AGFO), which activates rubric-based feedback rewards on demand during RL, reducing evaluation overhead while improving feedback quality. We also propose Adjacent Contrastive Reasoning (ACR) to improve ordinal score calibration by explicitly contrasting adjacent score levels. Experimental results show that the RFE framework captures essay-feedback consistency, exhibits strong pairwise discriminative power, and closely aligns with expert preferences. On the ASAP benchmark, RLAES-AGFO achieves the best scoring performance among LLM-based methods (QWK = 0.803), while maintaining feedback quality comparable to GPT-5.5 and avoiding the feedback degradation observed under score-only RL. Code and datasets are publicly available at https://github.com/hellomuyi/RLAES.

Abstract

Automated Essay Scoring

Explore similar work

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

PsyScore: A Psychometrically-Aware Framework for Trait-Adaptive Essay Scoring and ZPD-Scaffolded Feedback

When Rubrics Change: Cross-Rubric Generalization for Critical Thinking Essay Scoring