Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·10d ago·5 min read
ContractRL: Shielded Group-Relative Policy Optimization for Auditable Tool-Call Repair
TL;DR
ContractRL introduces a novel approach for repairing structured tool calls by using a contract-constrained sequential repair protocol. It achieves a semantic success rate of 0.9362 with fewer generated tokens compared to existing methods.
✦ Why It Matters
Implement ContractRL for more efficient and auditable tool-call repairs in your applications today.
Key Takeaways
- Utilize ContractRL for structured tool call repairs.
- Leverage contract-derived action masks to filter invalid operations.
- Optimize policy for improved semantic success rates.
- Analyze feedback mechanisms to enhance repair processes.
How It Works
ContractRL operates by modeling the repair of JSON structures as a sequential decision-making process. It uses a typed verifier to provide feedback on the candidate input, which, combined with an immutable repair history and a budget, informs the next action.
The action mask derived from the contract ensures that only valid operations are considered, thus streamlining the repair process and enhancing auditability.