Harvey LAB

Same capability. A shorter route through the work.

Output length per task across three model statesThree overlapping distributions: base with mean 22k, Standard RL with mean 90k, and density-aware training with mean 37k output tokens. Both trained variants have 8.3% all-pass rate. Base Standard RL Density-aware · mean 37k
Standard RLDensity-aware
BaseOpen weightsRun AStandard RLRun BDensity-aware
Mean tokens22k90k37k
All-pass rate0%8.3%8.3%

59% fewer output tokens.
The same all-pass rate.

Nemotron 3.5 Nano 30B-A3B. Curves traced from the reported figure ↗. The slider blends the two trained distributions; the table shows measured results. Density on the vertical axis means distribution density.