Evaluating Overlap Handling in Gemini 3.1 Flash Live: A Reproduction of Full-Duplex-Bench v1.5

We reproduce the Full-Duplex-Bench v1.5 evaluation pipeline [Lin et al., 2025] for Gemini 3.1 Flash Live Preview on Apple Silicon (M3 Max, no CUDA). In this paper, we extend the original benchmark by running four simulated overlap scenarios — background speech, talking to another person, user backchanneling, and user interruption — and evaluates model behavior using categorical dialogue labels, prosodic adaptation metrics, and stop/response latency. The findings show that Gemini responds correctly to user interruption in 72% of cases and resumes speech after background noise and side conversation in 50–63% of cases. We also found that pitch and speaking rate increase after overlap events across all tasks, while intensity decreases; this indicates that the model adapts to the speaker’s rhythm but does not assert dominance. Stop latency is lowest for user interruption (0.72s), where fast yielding is the correct behavior. The code and data are available at https://github.com/ifeanyidike/full-duplex-bench-repro.