LLM Reliability & Fairness in Automated Essay Scoring

When two LLMs score the same essay and disagree by over 2 points on a 32-point rubric, students graded by the stricter model face a structural disadvantage — this study measures exactly that.

---

📊 Read the Full Interactive Report

→ Open the scrollytelling report

Charts, scroll-driven narrative, and all findings in one interactive page. Works on mobile and desktop.

---

Abstract

The deployment of Large Language Models (LLMs) in Automated Essay Scoring (AES) for academic evaluation poses algorithmic bias risks if implemented without human oversight. This study evaluates the Mixture of Experts (GPT OSS 120B) and Dense (Qwen3-32B) architectures using a multiple run scoring approach on 150 IELTS like essays. These specific models were selected to contrast the evaluative capabilities of a high parameter sparse architecture against a more compact dense network. The evaluation encompasses severity bias detection, topic bias through unimodal stress testing, and a Human in the Loop (HITL) routing simulation. Results indicate a fundamental architectural trade off: Qwen3-32B achieved perfect format compliance but demonstrated score instability (ICC 0.84) and excessive leniency, whereas GPT OSS 120B was highly reliable (ICC 0.94) with precise discriminative power but suffered a 25.4 percent parsing failure rate due to format hallucination. Furthermore, both models exhibit conservatism bias on visual description tasks lacking image context. The variability based HITL simulation (CV < 15%, Range $\le$ 2) demonstrates that GPT OSS 120B can automatically approve 55.4 percent of essays, while flagging the remaining 44.6 percent for human review. These findings confirm that fully automated formative assessment presents considerable reliability challenges, establishing the HITL framework as an essential bridging solution to mitigate LLM bias and maintain academic integrity.

---

Research Summary

Item	Detail
Dataset	50 respondents × 3 essay prompts (150 theoretical max)
Evaluation runs	5 runs per essay per model
Rubric	4 criteria: Grammar, Lexical, Coherence, Task (max 32 pts)
GPT submitted / valid	130 submitted / 97 valid scored (33 JSON parse failures, 25.4%)
Qwen submitted / valid	128 submitted / 128 valid scored (0 failures)
Valid pairs for comparison	95 essays scored validly by both models

---

Key Findings

Metric	GPT-oss-120b	Qwen3-32b	Advantage
ICC (reliability)	0.943 — Excellent	0.844 — Good	GPT
Mean score / 32	25.47	28.32	Qwen
Consistency (mean CV)	0.047	0.092	GPT
Mean within-essay range	2.4 pts	5.56 pts	GPT
Auto-approve rate	55.4%	7.8%	GPT
JSON parse failure rate	25.4% (33/130)	0%	Qwen

• GPT is the more reliable scorer. ICC 0.943 (excellent) vs 0.844 (good). Score the same essay five times with GPT and the results stay within a 2.4-point range on average — less than half of Qwen's 5.56-point spread.
• Qwen scores higher but far less consistently. Mean score 28.32 vs 25.47, but 92.2% of its essays were flagged for human review due to high variance. GPT required review for only 44.6%.
• The scoring gap is a fairness risk. In 74 out of 95 paired essays, Qwen scored higher than GPT. The ~2-point systematic gap means a student graded by GPT is structurally disadvantaged compared to one graded by Qwen on the same essay.
• Qwen never fails to return a usable score. GPT failed to produce parseable JSON on 25.4% of essays under zero-shot conditions — a significant deployment blocker.
• Both models score data-report essays most strictly. Without visual context, neither model hallucinates chart content; both apply conservative scoring instead (GPT: 24.20, Qwen: 27.37).

---

Statistical Results

Severity Bias — Paired t-test

Computed on 95 valid paired essays (essays scored by both models with no parse failure).

Statistic	Value
t-statistic	−7.43
p-value	5.03 × 10⁻¹¹
GPT mean total	25.47 (SD = 7.02)
Qwen mean total	28.32 (SD = 6.00)

The gap is not random — it is systematic. With p = 5.03 × 10⁻¹¹, the probability this difference occurred by chance is effectively zero. GPT applies a consistently stricter standard than Qwen across all essays.

Individual Agreement — Bland-Altman Analysis

Computed on 95 valid paired essays.

Metric	Value
Mean difference (GPT − Qwen)	−2.07
Lower limit of agreement	−7.38
Upper limit of agreement	+3.25

The limits of agreement span 10.6 points, indicating wide individual-level disagreement beyond the systematic gap. The asymmetric bounds (wider on the negative side) reflect GPT's tendency toward occasional extreme strictness rather than symmetric noise.

Reliability — Intraclass Correlation (ICC-3)

5 runs per essay, two-way mixed consistency model. ICC measures how much of the score variation is signal (real essay quality) vs. noise (model inconsistency). A score above 0.9 is considered excellent.

Model	ICC	n essays	n balanced
GPT-oss-120b	0.943	97	81
Qwen3-32b	0.844	128	123

HITL Routing Simulation

Threshold: CV < 0.15 AND score range ≤ 2 → Auto-Approve; else Flagged for human Review.

Model	Auto-Approve	Flagged for Review
GPT-oss-120b	72 / 130 (55.4%)	58 / 130 (44.6%)
Qwen3-32b	10 / 128 (7.8%)	118 / 128 (92.2%)

Qwen's inconsistency effectively eliminates it from any pipeline that uses automated approval — 9 in 10 of its essays must go to a human reviewer, making it operationally equivalent to manual grading.

Topic Bias

Prompt Category	GPT mean	Qwen mean	Gap
data_report	24.20	27.37	3.17
social_policy_opinion	26.22	28.97	2.75
tech_society_opinion	27.18	28.85	1.67

The severity gap is widest on data_report tasks (3.17 pts) and narrowest on tech_society_opinion (1.67 pts), suggesting topic framing and visual context affect model scoring behavior differently.

---

Limitations

• Sample size: 150 essays from 50 respondents — moderate for an exploratory study, but findings should be replicated on a larger, more diverse corpus before drawing strong policy conclusions.
• Essay type: All prompts are IELTS/TOEFL-style academic writing. Results may not generalize to other writing contexts (creative writing, business reports, school essays).
• Effective GPT sample: GPT's 25.4% JSON failure rate reduces its usable sample from 130 to 97 essays, potentially introducing selection bias in ICC calculations.
• Zero-shot prompting only: Both models were tested without examples or fine-tuning. Providing few-shot examples or task-specific prompting may significantly change reliability and consistency results.
• Models tested: GPT-oss-120b and Qwen3-32b at time of study. Model behavior evolves with version updates.

---

Replication

To independently verify the reported numbers:

pip install pandas numpy
python scripts/extract_metrics.py

This reads directly from the evaluation JSONL files and prints total observations, paired intersection, and Bland-Altman metrics — matching data/processed/analysis/report.json.

---

Setup & Pipeline

Prerequisites

• Python 3.10+
• See requirements.txt for dependencies

pip install -r requirements.txt

Running the Pipeline

Three stages: preprocess → evaluate → analyze

Stage	Command
Preprocess	python src/preprocess.py --input data/raw/essays.xlsx --out-dir data/processed
Evaluate	python src/evaluate.py --inputs data/processed --models qwen3-32b gpt-oss-120b
Analyze	python src/analyze.py --inputs data/processed --out-dir data/processed/analysis

Run all three in sequence to generate reports.

Outputs

Located in --out-dir (default: data/processed/analysis):

File	Description
report.html	Interactive dashboard with KPI cards and embedded charts
report.md	Markdown summary with all tables
report.json	Machine-readable metrics
metrics_summary.csv	Per-model aggregated metrics
hitl_routing.csv	Per-essay HITL routing decisions
image/*.png	Static chart exports (Bland-Altman, heatmaps, boxplot, HITL)

Configuration

Edit constants in src/analyze.py:

• HITL_CV_THRESHOLD = 0.15 — auto-approve if CV below this
• HITL_RANGE_THRESHOLD = 2 — auto-approve if score range ≤ this

The analyze stage caches results. To force recompute:

python src/analyze.py --inputs data/processed --out-dir data/processed/analysis --force

---

Project Structure

llm-awe-reliability-fairness/
├── data/
│   └── processed/
│       ├── analysis/       # Generated reports and charts
│       ├── eval/           # Per-question model evaluation outputs
│       └── samples/        # Preprocessed essay samples (JSONL)
├── dataset/                # Source dataset (Excel)
├── notebook/               # Exploratory analysis notebook
├── scripts/
│   ├── extract_metrics.py  # Standalone metric extraction for reproducibility
│   ├── diag_checkpoint.py
│   ├── inspect_icc.py
│   └── inspect_xlsx.py
├── src/
│   ├── analyze.py          # Main analysis pipeline
│   ├── evaluate.py         # Model evaluation runner
│   ├── preprocess.py       # Data preprocessing
│   └── cli.py
├── scrolly-report.html     # Interactive scrollytelling report (live at GitHub Pages)
├── requirements.txt
└── LICENSE

---

Author

Muhammad Abdiel Al Hafiz GitHub: dlzcods

Citation

If you use this data or methodology in your work:

Al Hafiz, M. A. (2025). LLM Reliability & Fairness in Automated Essay Scoring:
Comparing GPT-oss-120b and Qwen3-32b on IELTS-style Writing Evaluation.
https://github.com/dlzcods/llm-awe-reliability-fairness

---

License

See LICENSE for details.