AdaMAST: Adaptive Failure Taxonomies for Enhancing LLM Agents
AdaMAST is a taxonomy system that analyzes and improves LLM agent failures.
AdaMAST is a taxonomy system designed to understand the failures of LLM agents. By utilizing failure codes derived from the behavior of an LLM, it identifies intervention points, thereby enhancing the performance of LLMs. This system leverages the agent's own traces to uncover previously undefined failure modes, enabling a continuous improvement process.