ContractRL Reduces Tool-Call Repair Tokens to 34.4 with Schema Constraints
October 2, 2026
ContractRL uses group-relative policy optimization to repair malformed JSON tool calls via a bounded decision process. By using contract-derived action masks and typed verifier feedback, the method achieves 0.9362 semantic success using only 34.4 generated tokens per repair.
HOW THIS AFFECTS YOU
●
builderYou can implement more efficient, lower-latency error recovery for agentic tool-calling workflows.
●
researcherThis approach offers a new framework for modeling verifier-guided repair as a constrained decision process.