A new paper on arXiv introduces Temporal Evidence Decomposition (TED), a training-free diagnostic for audio-visual large language models (AV-LLMs). The authors report that these models can fail at full-video speech-vision decisions even when useful visual evidence is locally recoverable from the video.
The finding suggests a disconnect between what the model can access in isolated frames and how it integrates that information across time. TED is designed to decompose temporal evidence, allowing researchers to identify where the reasoning chain breaks down without retraining the model.
Because the abstract is the only source available, details on the exact experiments and benchmark results are not yet public. The work points to a broader limitation in current AV-LLMs: local visual availability is not sufficient for global temporal composition.