Spatial
Masks and boxes in images.
A single region–text–region cycle extends naturally from images to videos.
Overview
Fine-grained intelligence needs both directions.
A model should not only describe a designated region, but also recover that region from language. Generalized CycleGRPO couples these dual capabilities with a geometrically verifiable reward, requiring neither caption annotations nor an external language-model judge during reinforcement learning.
Temporal
Target intervals and recurring events.
Spatio-temporal
Timestamped boxes and sparse trajectories.
Method
Actor as its own critic.
01
Describe
The policy generates candidate captions conditioned on a target region.
02
Re-localize
The same policy localizes each caption back into the original region space.
03
Learn
Overlap with the input region yields direct, verifiable rewards for both interfaces.
Results
Consistent gains, measured in each native space.
We compare matched base policies within each domain rather than mixing incompatible metrics or model families.
New benchmark
DOVE tests whether a caption identifies the right occurrence.
Distinctive Occurrence-level Video-caption Evaluation groups a target interval with semantically similar recurrences. Models view the complete video and describe a specified interval; a model-blind judge then answers verified temporal probes using only the generated caption.
Visual analysis
Descriptions remain tied to the designated evidence.
