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¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2 · i3 large
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i3 With the pooled marks adapter, every fine-tuned model grades as well as a human or better on both exams; the pooled adapter is significantly better in 9 of 10 model×exam cells (d from −0.94 to −1.13 on the ML exam). → Habibullah 2026 q2 i? Persona sensitivity nearly vanishes: pooled adapters move at most +0.32 MAE (marks recipe) across three harsh personas on both exams, where base models move up to 20 MAE points. → Habibullah 2026 (2)

Evidence¶

Habibullah 2026¶

Habibullah, A., Alshoibi, Y., Alshiekh, M., Khan, S., & Khan, N. (2026). Where LLM graders succeed and break: Evidence from two computer-science exams. arXiv. https://arxiv.org/abs/2609.29333 doi:10.48550/arxiv.2609.29333

q2 · i3 · causal · r2

LoRA fine-tuning of five open models (4B–30B) on seed-fixed 80/20 student splits, evaluated on held-out students with a paired third-grader test. The printed paired effect sizes are large (d from −0.94 to −1.13 on the ML exam); the pooled fine-tune "lands at 1.75–2.01 MAE on the CV exam and 3.29–3.53 on the ML exam (marks recipe; Table 1)".

"The pooled marks adapter is significantlybetterin 9 of 10 (model× exam) cells: all five models on the ML exam (d from−0.94 to−1.13, every CI below zero) and four of five on the CV ."

Habibullah 2026 (2)¶

Habibullah, A., Alshoibi, Y., Alshiekh, M., Khan, S., & Khan, N. (2026). Where LLM graders succeed and break: Evidence from two computer-science exams. arXiv. https://arxiv.org/abs/2609.29333. doi:10.48550/arxiv.2609.29333

q2 · i? · causal · r2

Persona sweep re-run for base and pooled models on both exams. Bases reproduce the collapse (Qwen2.5-Coder-14B goes 4.06→23.81 under strict on the CV exam), while under a neutral prompt the worst base mis-grades 43 of 114 CV-exam students by more than 10 marks and every pooled model is at 0–1.

"The pooled adapters are flat everywhere: at most +0.32 (marks) and +0.62 (bd) across three personas, two exams and five models, and the catastrophic tail the collapse creates is gone:"

Discussion¶

Related Claims¶

• Grading skill transfers across exams: an adapter fine-tuned on one exam improves grading on the unseen other, though transfer stops short of in-domain performance — related
• A short strict-grader preamble knocks 14 of 17 open-weights LLM graders out of the graded band on a computer-vision exam, with three stopping grading outright — related
• Six closed-model configurations grade at or below the human inter-grader floor of 2.61/35 on the CV exam, with the best reaching MAE 1.64 — related
• The strict-persona vulnerability replicates on an independent machine-learning exam, but its direction is exam-specific: seven of 17 models improve because the persona cancels their over-marking — related
• The strict-persona collapse is attributed to the preamble's two credit-withholding policy sentences, not its harsh tone, with S2 alone causing refusal in two models — related