When VLMs'Fix'Students: Identifying and Penalizing Over-Correction in the Evaluation of Multi-line Handwritten Math OCR

2026-04-01

AI Summary

Existing evaluation methods for handwritten mathematical OCR struggle to detect how vision-language models (VLMs) often obscure genuine student errors in multi-line solutions through “over-correction.” This work presents the first systematic investigation of this issue and introduces PINK, a novel evaluation framework grounded in semantic faithfulness. PINK integrates large language model–guided scoring rules, semantic alignment analysis, and INK format parsing to explicitly penalize unfaithful transcriptions. Evaluations of 15 state-of-the-art VLMs on the FERMAT dataset reveal that PINK yields rankings substantially divergent from those based on traditional BLEU scores and aligns more closely with human expert judgments (55.0% vs. 39.5% preference), demonstrating its superior suitability for educational assessment scenarios.

Abstract

Accurate transcription of handwritten mathematics is crucial for educational AI systems, yet current benchmarks fail to evaluate this capability properly. Most prior studies focus on single-line expressions and rely on lexical metrics such as BLEU, which fail to assess the semantic reasoning across multi-line student solutions. In this paper, we present the first systematic study of multi-line handwritten math Optical Character Recognition (OCR), revealing a critical failure mode of Vision-Language Models (VLMs): over-correction. Instead of faithfully transcribing a student's work, these models often"fix"errors, thereby hiding the very mistakes an educational assessment aims to detect. To address this, we propose PINK (Penalized INK-based score), a semantic evaluation metric that leverages a Large Language Model (LLM) for rubric-based grading and explicitly penalizes over-correction. Our comprehensive evaluation of 15 state-of-the-art VLMs on the FERMAT dataset reveals substantial ranking reversals compared to BLEU: models like GPT-4o are heavily penalized for aggressive over-correction, whereas Gemini 2.5 Flash emerges as the most faithful transcriber. Furthermore, human expert studies show that PINK aligns significantly better with human judgment (55.0% preference over BLEU's 39.5%), providing a more reliable evaluation framework for handwritten math OCR in educational settings.

Problem

• handwritten math OCR
• over-correction
• Vision-Language Models
• educational assessment
• multi-line expressions

Innovation

• over-correction
• handwritten math OCR
• semantic evaluation metric
• PINK
• Vision-Language Models