An experimental video AI system reported a 79-fold reduction in the time needed to complete a spatial tube, while increasing spatial decoding throughput 92-fold on the VidSTG benchmark. It also improved grounding accuracy against standard autoregressive decoding, which generates outputs step by step.
The approach, called Parallel Tube Decoding, or PTD, changes how the model spends its sequential work. The benchmark formulation asks for one continuous temporal interval and one spatial tube per query. PTD removes token-level and trajectory-level dependencies, using one temporal decoding round followed by one spatial decoding round regardless of the tube's length.
The model trims the sequential work
Within that schedule, PTD predicts a temporal block, then time-conditioned spatial blocks in parallel, rather than making each box depend on the previous one. It uses Decoupled Block Attention to preserve shared video-query context while removing cross-box dependencies. The method also adds localization-aware policy optimization for temporal boundaries and spatial geometry.
The model uses a Qwen3-VL-4B backbone with discrete spatial and temporal localization tokens. Training had two stages: supervised fine-tuning with the dual NTP/MTP formulation, followed by localization-aware Group Relative Policy Optimization, or GRPO. The setup used LoRA with rank 32, and the combined VidSTG and HC-STVG training data yielded about 90,000 SFT samples and a 16,000-sample GRPO training set.
The study uses two speed measures. Tube Completion Latency, or TCL, measures tube-completion time after the first token. Boxes Per Second, or BPS, measures spatial decoding throughput. In plain terms, TCL tracks the wait after generation has started, while BPS shows how quickly the spatial boxes are produced.
Speed held as the tube grew
Among the decoding options compared, PTD had the strongest listed grounding performance, with a TCL of 0.4 seconds and BPS of 45.9 boxes per second. The study reports no confidence intervals or significance tests for these comparisons.
The advantage became clearer as the requested tube grew. When the number of boxes rose from 8 to 64, PTD latency moved only from 0.33 to 0.40 seconds. Under block decoding, latency rose from 0.72 to 6.18 seconds over the same range. This scaling analysis is descriptive and includes no variability estimates.
Accuracy gains traveled beyond one benchmark
Accuracy gains appeared on the main grounding benchmarks as well. On VidSTG, PTD led on seven of eight reported metrics. Its mean temporal IoU, or mtIoU, improved on the previous best by 1.1 points for declarative queries and 1.7 points for interrogative queries. Mean video IoU, or mvIoU, improved by 1.1 points for both query types.
On HC-STVG-v1 and HC-STVG-v2, the model improved the previous best mvIoU by 1.5 and 1.1 points, respectively. It also achieved the best video IoU at the 0.3 threshold on both benchmarks.
The reported gains extended to tests outside the core spatio-temporal grounding benchmarks. In zero-shot evaluation, PTD improved mean performance on Charades-STA by 4.7 points, ActivityNet mIoU by 9.1 points and ReXTime mIoU by 15.6 points. On Ref-DAVIS, Ref-YT-VOS and ReasonVOS, it exceeded Molmo2-4B by 8.4, 2.6 and 0.2 points, respectively, when the systems used the same SAM2 segmentation backend.
That tracking comparison needs a qualification. Because those results used an off-the-shelf segmentation model, the paper does not show that the tracking gains came from PTD alone.
The hard cases are still open
The method's ablation points to the value of combining both kinds of feedback. The combined temporal-plus-spatial reward produced the strongest reported VidSTG results among the tested reward variants: declarative mtIoU was 53.7 and mvIoU 38.3, while interrogative mtIoU was 52.2 and mvIoU 32.4.
There are clear boundaries to what the results cover. The benchmark formulation uses one continuous temporal interval and one spatial tube per query, leaving disjoint events, periods when a target moves out of view and multiple instances matching the same expression largely unexplored. The model was also tested with a single Qwen3-VL-4B backbone, so the findings do not establish performance across other backbone sizes or architectures.
Spatial localization was harder after occlusion and for small, rapidly moving targets, while temporal failures clustered around brief state changes whose boundaries were visually ambiguous. The paper is an arXiv version 1 preprint dated 28 August 2026.
Paper data and sources
Original title: Locate Anything in Videos: Rethinking Efficient Generative Spatio-Temporal Video Grounding
Authors: Hanoona Rasheed, Haania Siddiqui, Ming-Hsuan Yang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text