Event
TRIAGE: VLA Inference Reuse and Pruned Visual Tokens
1 reports · 1 sources
Sooyoung Kwon, Carlee Joe-Wong, and Kyunghan Lee propose TRIAGE, a training-free acceleration method for vision-language-action (VLA) model inference: it uses a token drift threshold to determine whether cached KVs remain valid, reusing them if valid and pruning them otherwise, and fixes a recomputation budget to directly control prefill latency. The authors claim that on LIBERO-Object, when recomputing only 7% of visual tokens, the method outperforms the strongest baseline by 24.4 percentage points; the same diagnostic method can calibrate this threshold separately for different backbone networks.
Report timeline
- TRIAGE: VLA inference acceleration requires both reusing and pruning visual tokens
TRIAGE proposes a training-free visual token acceleration method that decides whether to reuse cached KV or prune based on a token drift threshold, and fixes the recomputation budget to directly control prefill latency.
arXiv cs.RO ·