HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention

Researchers have introduced HERA (Harness-Environment Co-Evolution), a method for improving agentic abstention — the ability of large language model agents to recognize when a task is infeasible and no valid solution exists. HERA addresses a reliability gap in LLM agents that act in complex tool-use environments, where agents often fail to abstain from tasks that cannot be completed. The work builds on recent formalization of agentic abstention as a distinct problem, targeting existing approaches' shortcomings.

Large language model LLM agents are increasingly capable of acting in complex tool-use environments, yet they often fail to recognize when tasks are infeasible and no valid solution exists. Recent work has formalized this reliability gap as the problem of agentic abstention, and existing approache