Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme

A recent research paper highlights a critical discrepancy in evaluating performance, particularly for language models, when traditional accuracy metrics are applied to assessment schemes that utilize non-additive, convex grading. The 2025 reform of Vietnam's National High School Graduation Examination serves as a case study, demonstrating how a convex grading system for multi-part questions penalizes partial knowledge more severely than standard proportional credit, leading to an underestimation of 'correctness' by conventional metrics. This has implications for how performance is measured and interpreted in contexts with complex scoring.

Why it matters

This analysis underscores the importance of aligning evaluation methodologies with the specific scoring mechanisms of an assessment, particularly in high-stakes environments. Misalignment can lead to inaccurate performance measurement, potentially misrepresenting capabilities or outcomes, and influencing resource allocation or strategic direction based on flawed data. Organizations need to understand the nuances of grading to ensure fair and accurate assessment, whether for human or automated systems.

Key insights

• Traditional evaluation benchmarks often assume proportional credit for partial knowledge, scoring responses as simply right or wrong.
• Non-additive, convex grading schemes invalidate the assumption that partial knowledge is worth proportional credit.
• Vietnam's 2025 National High School Graduation Examination Part II uses a convex grading scheme for questions with four true/false statements.
• In this scheme, three correct statements out of four earn 0.50 points, not the 0.75 points expected from standard proportional accuracy metrics.
• This difference illustrates a 'partial-credit gap' where partial knowledge is awarded less than proportional credit.
• Part II of the exam contributes significantly, accounting for 4.0 points (implying 40% of a 10-point scale, though the full scale is not specified).

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.18336