CodeAssay is a 185-task Python benchmark showing that auditing benchmark ground truth flips 9% of model correctness labels and widens measured model spread, and that a security-focused prompt costs code size and complexity without measurable security gains.
Biometrika 37(3–4), 256–266 (1950) 16 S
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SE 1years
2026 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
CodeAssay: A Multi-Metric Benchmark with Audited Ground Truth for LLM Code Generation
CodeAssay is a 185-task Python benchmark showing that auditing benchmark ground truth flips 9% of model correctness labels and widens measured model spread, and that a security-focused prompt costs code size and complexity without measurable security gains.