IHBench Evaluates Voice Agent Recovery from User Interruptions

Voice agents operating in structured environments, such as customer service or healthcare scheduling, frequently encounter user interruptions. While existing benchmarks focus on the timing of these interruptions, they often overlook the crucial aspect of what happens *after* an interruption: whether the agent can correctly resume the workflow, address the interjection, and avoid repeating information. To fill this gap, researchers have developed IHBench, the Interruption Handling Benchmark. This benchmark specifically evaluates voice agents' ability to recover from interruptions while executing state-machine-driven workflows across ten diverse enterprise domains. It injects six different types of interruptions at controlled points during an utterance, with each interruption scored on both task fulfillment and the quality of the recovery. An evaluation of 27 audio-language model configurations from major providers like OpenAI and Google, as well as open-weight models, revealed significant performance disparities. Closed-weight models consistently demonstrated greater robustness to interruptions, outperforming open-weight models in task fulfillment, degrading slower with conversation length, and showing no modality gap between audio and text inputs. A human study validated the LLM judge's accuracy, confirming that recovery quality is a distinct and important capability for voice agents.