Benchmark
MRBench
Assess how tutors respond to student mistakes and confusion in mathematics.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
192 conversations and 1,596 responses from human and model tutors, annotated across eight pedagogical dimensions.
Methodology
Human annotations evaluate the pedagogical value of responses. The study also checks model-based evaluators against those annotations.