March 2026 Summaries

Voice agents are evolving from traditional turn-based designs to more natural, continuous interaction systems, reflecting a shift toward real-time human-like communication. Modern full-duplex spoken dialogue systems, such as GPT-realtime-2025-08-28 and Gemini Live-2.5-flash-native-audio, handle speech continuously, accommodating interruptions and mid-utterance updates, unlike older models limited by turn boundaries. EchoChain, a new benchmarking tool, evaluates Dual-Stream Reasoning (DSR) in these systems by simulating multi-turn conversations with context-grounded interruptions, testing models on their ability to integrate new information seamlessly. Despite advancements, current models often struggle with interruption-induced errors like Contextual Inertia, Interruption Amnesia, and Objective Displacement, where they fail to maintain task coherence under changing inputs, as evidenced by a top-performing model achieving only a 47.5% success rate. EchoChain provides a structured framework for assessing these models' real-time reasoning capabilities, revealing significant room for improvement in managing dynamic interactions, which is critical as voice assistants become more prevalent in everyday use.