In the authors' practical test, vision-capable LLMs recognized handwritten examination letters with high accuracy, with Gemini 3 Flash best at about 98.84% (35 errors in 3009 letters)¶

Claim · All claims
Evidence · 1 study · design 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? Across 3009 letters recognized in practical use, Gemini 3 Flash produced 35 incorrect recognitions requiring manual post-processing, about 98.84% accuracy, outperforming Qwen3-VL-30B-A3B-Instruct (~98.31%), GPT-5 (~96.54%), and Grok 4.1 Fast (~88.90%). → Hartwig Grabowski 2026

Evidence¶

Hartwig Grabowski 2026¶

Hartwig Grabowski, Michael Canz. (2026). Hybrid E-Assessment in Higher Education: Semi-Automated Grading of Paper-Based Written Examinations. https://arxiv.org/abs/2606.08855 doi:10.48550/arxiv.2606.08855

q2 · i? · design · r2

Practical evaluation of vision-capable LLMs on scanned examination answer tables: "Across 3009 letters to be recognized, 35 incorrect recognitions occurred and required manual post-processing". Table 1 prints error rates for four models; Gemini 3 Flash showed the best performance in the tested setting. No standardized effect size is printed.

"Google Gemini 3 Flash showed the best evaluation performance in the tested setting. Across 3009 letters to be recognized, 35 incorrect recognitions occurred and required manual post-processing (Table 1)."

Discussion¶

Related Claims¶

• An earlier YOLO-5-based handwriting recognition method reached only 88.28% accuracy on examination table fields, which the authors judge insufficient for summative assessment — related
• Most AI grading errors (87% for the best model) were attributable to transcription failures rather than rubric misapplication