T2J-Bench tests coding agents on 63 whole-repository PyTorch-to-JAX migrations

A researcher sharing work at Google introduces T2J-Bench, with 63 whole-repository migration tasks across 18 model families covering foundation-model training and serving. They say the strongest off-the-shelf configuration reaches 0.676 reward against 0.980 for a reference JAX oracle. The team also introduces TwinHarness, a migration loop they say raises reward from 0.649 to 0.723 with the same agent and model—an 11.4% relative gain.