SAPO
A self-supervised reinforcement learning method that requires no manual temporal annotations. It uses the internal temporal attention of a VLM as a pseudo-label signal to drive GRPO optimization of video temporal grounding capability.
Timeline (last 90 days)
The core of Semantic-Aware Policy Optimization (SAPO) is a semantic-aware gating mechanism that constrains TD residuals from failed atomic subtasks, preventing them from propagating across task boundaries.
AtomicRL's average 21.3% real-world improvement depends on the effectiveness of the Real2Sim2Real transfer pipeline, whereas the improvements on simulation benchmarks mainly come from SAPO and BATE's improvements to credit assignment and exploration, so the two have different sources of contribution.
SAPO achieves a 33.7% relative mIoU gain on Qwen2.5-VL-3B.
SAPO achieves a 5.4% relative mIoU gain on Time-R1-7B.
SAPO (Self-supervised Attention Policy Optimization) uses the model's internal temporal attention as a self-supervised signal to drive Group Relative Policy Optimization (GRPO) to improve video temporal grounding.
The core idea of SAPO is to bypass reliance on manual annotation and optimize directly against the temporal generation gap, enabling the model to more accurately express information it already knows internally.
In the future, ideas similar to SAPO may be extended to other multimodal tasks that require precise output, such as spatial grounding or fine-grained action recognition.
SAPO's improvements transfer across datasets, and while improving grounding accuracy, the model's performance on general video understanding tasks does not degrade.
SAPO uses the attention interval as a pseudo-label to evaluate prediction quality and designs screening criteria that require no ground-truth annotations to pick out high-value training samples.
All Facts (9)
AtomicRL's average 21.3% real-world improvement depends on the effectiveness of the Real2Sim2Real transfer pipeline, whereas the improvements on simulation benchmarks mainly come from SAPO and BATE's improvements to credit assignment and exploration, so the two have different sources of contribution.
50%UnverifiedThe core of Semantic-Aware Policy Optimization (SAPO) is a semantic-aware gating mechanism that constrains TD residuals from failed atomic subtasks, preventing them from propagating across task boundaries.
50%UnverifiedSAPO achieves a 33.7% relative mIoU gain on Qwen2.5-VL-3B.
50%UnverifiedSAPO achieves a 5.4% relative mIoU gain on Time-R1-7B.
50%UnverifiedSAPO (Self-supervised Attention Policy Optimization) uses the model's internal temporal attention as a self-supervised signal to drive Group Relative Policy Optimization (GRPO) to improve video temporal grounding.
50%UnverifiedThe core idea of SAPO is to bypass reliance on manual annotation and optimize directly against the temporal generation gap, enabling the model to more accurately express information it already knows internally.
50%UnverifiedSAPO's improvements transfer across datasets, and while improving grounding accuracy, the model's performance on general video understanding tasks does not degrade.
50%UnverifiedSAPO uses the attention interval as a pseudo-label to evaluate prediction quality and designs screening criteria that require no ground-truth annotations to pick out high-value training samples.
50%UnverifiedIn the future, ideas similar to SAPO may be extended to other multimodal tasks that require precise output, such as spatial grounding or fine-grained action recognition.
50%