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¶

Claim · All claims
Evidence · 1 study · causal r2 · q2
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 i? Mean MAE on the ML exam falls from 7.95 (base) to 5.36 under the CV-exam adapter, and on the CV exam from 5.54 to 3.57 under the ML-exam adapter, with every model improving in both directions but one exception. → Habibullah 2026

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 · i? · causal · r2

Cross-exam transfer evaluation of per-exam LoRA adapters on held-out students of the unseen exam. The article notes the one negative transfer cell is Gemma-4-E4B on the ML exam (4.66→4.79), the strongest base there with the least room to improve.

"An adapter fine-tuned on one exam improves grading on the other, which it never saw: mean MAE on the ML examfalls from 7.95 (base) to 5.36 under the CV-exam adapter, and on the CV exam from5.54 to 3.57 under the ML-examadapter."

Discussion¶

Related Claims¶

• 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
• 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 — related
• LLMs are more stable than supervised models across dataset shifts, though they do not surpass within-dataset fine-tuned performance — a broader claim this one bears on
• Transfer between two specific datasets is asymmetric: Sangodiah-trained models transfer unusually well to Gani questions — 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
• Supervised ML/DL Bloom question classifiers trained on one dataset drop substantially in weighted F1 when tested on unseen datasets — related