Benchmark
M4Bench
Test whether coding agents reproduce clinical variables from electronic health records.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
28 derivation tasks across 15 families using MIMIC-IV and eICU. Contributor links show four of the eight paper authors.
Methodology
Compare generated derivations with reference checks. Controls separate specification detail, skill format and schema familiarity. Partial agreement does not imply every check passed. Schema perturbations are descriptive checks, not a bound on memorization. Paper v1 is not a benchmark edition.