Research Article · 2026

Can Human-in-the-Loop
Mitigate LLM Scoring Bias?

Automated IELTS-like essay scoring with GPT-oss-120b & Qwen3-32b, routed through a Human-in-the-Loop framework

Muhammad Abdiel Al Hafiz · llm-awe-reliability-fairness

0
Essays Collected
0 Runs
Per Essay Per Model
0 ICC
GPT Reliability Score
0%
Qwen Flagged for Review

Two Models, One Test: Can LLMs Grade Fairly?

We tested whether two large language models can score IELTS and TOEFL essays reliably, without systematic bias. If they can, automated scoring becomes a real option. If they cannot, any deployment carries hidden fairness risks.

GPT-oss-120b (Mixture-of-Experts) and Qwen3-32b (Dense) were each given the same 150 essay collection. GPT successfully completed scoring for 130 of them; Qwen for 128. Each completed essay received 5 independent scoring passes to measure self-consistency.

📋
Study design: 150 essays collected · GPT scored 130 (97 with valid JSON) · Qwen scored 128 (all valid) · 3 prompt types · 4 rubric criteria (Grammar, Lexical, Coherence, Task) · Max score: 32 points
Source: data/processed/analysis/report.json · hitl_routing.csv

Intraclass Correlation: How Repeatable Is Each Model?

ICC tells you whether a model gives the same essay the same score when you run it multiple times. A high ICC means the model is self-consistent. A low ICC means the scores shift too much to be trusted.

GPT-oss-120b reaches an ICC of 0.943 on its 97 valid essays, putting it in the Excellent range. Qwen3-32b scores 0.844 across all 128 valid essays, firmly Good but short of Excellent.

📊
ICC thresholds: below 0.50 Poor · 0.50 to 0.75 Moderate · 0.75 to 0.90 Good · above 0.90 Excellent
⚠️
A gap of 0.1 ICC points between models sounds small. In high-stakes assessment, where consistency is non-negotiable, it is not.

GPT Scores About 2 Points Below Qwen Per Essay

A paired t-test on 95 essays scored by both models finds an average gap of 2.85 points between the two overall means (t = 7.43, p < 0.001). The Bland-Altman analysis puts the per-essay mean difference at 2.07 points, with individual essays varying anywhere from 7.38 points below to 3.25 points above.

The Bland-Altman plot shows this clearly. If both models scored the same way, the dots would scatter evenly around zero. Most fall below, pointing to GPT consistently applying a stricter standard.

🚨
p = 5.03 × 10−11. The gap is real. Scoring essays from the same cohort with both models gives students graded by GPT a structural disadvantage.

Essay by Essay: Qwen Scores Higher in Almost Every Case

These are the actual scores for all 95 essays both models evaluated, sorted from lowest to highest. Each point on the blue line is GPT's score for that essay. Each point on the amber line is Qwen's.

The two lines move together — both models rank essays in roughly the same order — but the amber line stays above the blue almost all the way across. That gap is the 2-point severity bias made visible.

📉
In 74 out of 95 essays Qwen scored higher than GPT. In only 21 did GPT score higher. The gap is not random noise — it is structural.
Derived from 95 matched pairs (mean + diff/2 per model). Sorted by essay mean score ascending.

Run-to-Run Stability: GPT Is 2x More Consistent

Consistency measures how much a model's score changes when you run it again on the same essay. Two numbers capture this: Coefficient of Variation (CV) and score range across 5 runs.

GPT's mean CV is 0.047. Qwen's is 0.092, nearly twice as variable. GPT's average score range across runs is 2.4 points. Qwen's is 5.6 points.

GPT's CV of 0.047 means scores shift less than 5% across runs. You can trust a single GPT score. With Qwen at 0.092, averaging multiple runs is safer.

Prompt Category Matters: Data Tasks Get the Lowest Scores

Both models score essays differently depending on the prompt type. Essays that describe data, charts, or graphs receive the lowest scores from both models across the board.

The gap between models is widest for data report prompts at 3.17 points, and narrows for opinion tasks. Both models apply stricter criteria when the prompt expects students to interpret visual data that is absent from the text.

🗺️
Neither model fabricates visual content it cannot see. Both score more strictly when the expected chart or graph context is missing from the input.

Automation Suitability: GPT Clears 7x More Essays Without Human Review

Human review routing flags any essay where the model's 5 runs are too inconsistent: CV at 0.15 or above, or a score range above 2 points. Those essays go to a human grader before being used.

GPT passes 55.4% of essays straight through without review. Qwen flags 92.2% of essays, meaning a human needs to check almost every Qwen score before it can be used.

⚠️
JSON compliance: GPT fails to return valid JSON in 25.4% of runs (33 out of 130 essays). When parsing fails, the score cannot be used at all. Qwen returns valid JSON every time.

Three Takeaways for Automated Essay Scoring

The radar chart compares both models across five dimensions where higher is better. No model wins everywhere. Which one to use depends on what your pipeline needs most.

🏆
Use GPT when repeatability is the priority. An ICC of 0.943 and a CV of 0.047 mean GPT scores stay stable across runs, making it the safer choice for automated pipelines.
🚫
Never score the same cohort with both models. The 2-point systematic gap means students graded by GPT face a structural disadvantage compared to those graded by Qwen.
🛠️
Fix GPT's JSON output before going to production. A 25.4% parse failure rate means one in four essays gets no usable score. Qwen's 100% JSON compliance is a real engineering advantage.
ICC model: two-way mixed, consistency (pingouin). Paired t-test: scipy.stats. n = 95 matched pairs.
Study at a Glance: Two Models, Three Prompt Types
GPT-oss-120b (Mixture-of-Experts) vs Qwen3-32b (Dense) · 150 essays collected · 3 essay categories · 5 independent scoring runs each.
Intraclass Correlation Coefficient: ICC(2,k) Consistency Model
GPT-oss-120b
Qwen3-32b
Semi-circle gauges span 0–1. Threshold zones: 0.90 = Excellent (blue), 0.75 = Good (gold), 0.50 = Moderate, below = Poor.
Bland-Altman Plot: 95 Matched Essay Pairs
Mean difference (GPT − Qwen)
±1.96 SD limits of agreement
GPT scored higher
Qwen scored higher
X = mean of both scores. Y = GPT minus Qwen. Mean diff = −2.07 (red line). LoA: −7.38 to +3.25.
Score Movement: GPT vs Qwen Across 95 Essays
GPT-oss-120b
Qwen3-32b
95 essays scored by both models, sorted by mean score (low → high). Y axis = total score (max 32 points).
Consistency Metrics — CV & Score Range Across 5 Runs
GPT-oss-120b
Qwen3-32b
CV (Coefficient of Variation) = SD / Mean. Lower is better. Score range = max minus min across 5 runs per essay.
Mean Scores by Prompt Category & Model
Lower scores
Higher scores
Color scale: RdBu from 22 (red) to 30 (blue). Both models score data_report essays most strictly. Gap (Qwen−GPT) shown below.
HITL Routing Distribution & JSON Compliance
Auto-Approve (CV < 0.15 & range ≤ 2)
Flagged for Human Review
Auto-approve threshold: CV < 0.15 AND score range ≤ 2 points. JSON failure counts as flagged for GPT.
Multi-Dimensional Comparison: Five Key Axes
GPT-oss-120b
Qwen3-32b
Axes normalized to 0–1 (higher = better). Consistency = inverse CV. HITL Efficiency = auto-approve rate. JSON Compliance = 1 − parse-failure rate.