Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 26·5 min read
When May an Agent Stop? Evidence-Carrying Termination for Tool-Using LLMs
TL;DR
Evidence-Carrying Termination (ECT) was developed to enhance decision-making in tool-using agents, ensuring they only claim completion when supported by valid evidence. ECT demonstrated zero unsafe completions across 288 tasks, significantly outperforming existing termination methods.
✦ Why It Matters
Implement ECT in your tool-using agents to ensure safer and more reliable completion claims.
Key Takeaways
- Adopt Evidence-Carrying Termination for safer agent operations.
- Ensure agents provide evidence for all claims made.
- Evaluate agent performance using synthetic tasks for reliability.
- Implement recovery mechanisms to handle execution faults.
Full Summary
Tool-using agents, such as those powered by large language models (LLMs), face challenges in determining when to stop their operations. Existing methods often fail to adequately verify the correctness of their outputs.
The newly introduced Evidence-Carrying Termination (ECT) requires agents to provide a typed certificate that links each answer claim to valid evidence from their execution traces. In a study involving 48 synthetic tasks across six tool-use families, ECT achieved 0 unsafe completions compared to 252 unsafe completions from traditional methods.
Additionally, in a separate evaluation of 66 trajectories, ECT prevented premature terminations in 0 cases, while the previous method had 40. ECT not only ensures correctness but also allows for successful recovery in 18 out of 66 cases, demonstrating its robustness in handling faults.
This approach emphasizes the importance of evidence in certifying agent actions, moving beyond reliance on external truths.