Temporal Generation Gap
The phenomenon whereby the time intervals indicated by a VLM's internal attention mechanism are more accurate than the timestamps in its final output, indicating that the model loses precision when converting internal representations into concrete timestamps.
Timeline (last 90 days)
In about 66% of valid samples, the time interval indicated by the VLM's internal attention mechanism is more accurate than the timestamp output by the model's final prediction.
The research team analyzed the temporal generation gap phenomenon on four VLMs and two grounding datasets.
All Facts (2)
In about 66% of valid samples, the time interval indicated by the VLM's internal attention mechanism is more accurate than the timestamp output by the model's final prediction.
50%UnverifiedThe research team analyzed the temporal generation gap phenomenon on four VLMs and two grounding datasets.
50%