Validator agent reliability is highest for realism and lowest for authenticity in human evaluation¶

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? The realism validator agent matches human majority judgments with fair agreement (κ=0.322, acc=0.667), while the authenticity validator shows only slight agreement (κ=0.133, acc=0.511). → Fareya Ikram 2026

Evidence¶

Fareya Ikram 2026¶

Fareya Ikram, Nischal Ashok Kumar, Junyang Lu, Hunter McNichols, Candace Walkington, Neil Heffernan, and Andrew S. Lan. (2026). A Multi-Agent Approach to Validate and Refine LLM-Generated Personalized Math Problems. arXiv preprint. https://arxiv.org/abs/2604.05160 doi:10.48550/arxiv.2604.05160

q2 · i? · design · r2

Human evaluation in which three annotators with college-level mathematics background labeled 45 problems sampled from the Zero-Shot baseline outputs for authenticity, realism, and readability using the same rubric as the validator agents; solvability was excluded due to highly skewed labels. Cohen's κ and accuracy against the human majority label are reported in Table 3.

"For realism (κ=0.322, acc=0.667) and readability (κ=0.281, acc=0.800), there isfairagreement between validator agents and human annotators. For authenticity (κ=0.133, acc=0.511), there is onlyslightagreement."

Discussion¶

Related Claims¶

• Validator failure patterns vary by interest topic and math curriculum unit — related
• Authenticity and realism are the most frequent failure modes in initial LLM-personalized math problems, and a single refinement iteration substantially reduces failures — related
• Realism issues were detected by agents during generation but were rare in final versions — related
• Human-LLM agreement remains low across constructs (κ = 0.000 to 0.419), with self-efficacy showing the best alignment and Prior KSAs none — related