Put reasoning outside the weights

PLVR puts part of reasoning outside the base model. The method learns an explicit program of deterministic and neural primitives, gives each primitive a typed input/output contract, and uses symbolic backpropagation to send required output types backward through the program. The base model stays frozen while the program and its small primitive models are learned. 12

The product problem is inspectability. A terminal reward can tell a training loop that an answer passed; a typed contract can point to the step that failed and constrain the next search. On LiveCodeBench v6 and τ²-Bench, the paper reports average gains of 48.5 points for GPT-OSS-20B and 22.7 points for Muse Glimmer 30B, with a 27.8-point average advantage over matched-budget RL comparisons. 2

A first build needs one task with checkable intermediate results: define typed primitives, write validators, keep the current route beside the PLVR branch, and compare task success, contract failures, latency, and cost. The paper page leaves the promised library and conformance checker’s public locator unspecified. Treat the release path as an implementation sketch until the artifact is verified. 2