Videos are expensive to analyze frame by frame, yet many video understanding tasks depend on knowing where relevant moments occur. A system may need to find when an action changes, locate the segment described by a sentence, or choose a few frames for a vision-language model. Existing methods often solve these problems separately, using task-specific training data or specialized architectures. We study whether a pretrained video-text model can provide enough temporal structure to support several of these tasks at once. We present STITCH, a training-free method that divides a video into semantically meaningful temporal chunks. STITCH embeds short video windows with a frozen video-text backbone and detects changes in the resulting embedding sequence. These chunks are computed once per video and reused across tasks. We evaluate STITCH on generic event boundary detection, language-based moment retrieval, and frame selection for long-video VLM reasoning. Across all three settings, STITCH remains competitive with more specialized methods while requiring no task-specific training, with especially clear gains when only a small number of frames or tokens can be processed. These results suggest that reusable temporal abstraction is a promising direction for general video understanding, allowing dense video streams to be converted once into semantic units that can be localized, retrieved, sampled, or reasoned over by downstream systems.
Figures & tables
Figure 1: STITCH converts semantic change into a reusable video timeline. As described in Section 3 , STITCH first encodes short video windows with a frozen video-text backbone, semantic change points are then detected in embedding space, and the video is partitioned into adaptive chunks. The same chunked timeline can be interpreted directly as event boundaries, scored with a language query for retrieval, or used to select compact visual evidence for long-video reasoning.
Figure 2: GEBD : STITCH boundaries reveal interpretable event transitions. In the event detection setting of Section 5 , STITCH uses chunk boundaries from the shared temporal abstraction as event boundary predictions. The text labels are manually added descriptions of the visible events, not outputs of the model; they are included to clarify what the predicted cuts separate in the video.
Table 1: Cross-task summary of specialized baselines and STITCH. The table groups prior methods by their reported task scope, so the N/A blocks indicate tasks outside a method’s evaluated setting rather than missing results. STITCH is evaluated across all three task families using the same training-free temporal abstraction front end. For event detection and moment retrieval, we report the representative metrics shown in the column headers; for long-video VLM QA, Δ denotes the accuracy gain over uniform frame sampling. All values are percentages. VLM results use LLaVA-OV 7B [ 19 ] with eight input frames and the MMR-based STITCH selector. Full metrics, uniform baselines for the VLM deltas, and additional STITCH variants are reported in Section A . Unified localization frameworks are discussed in Section 2 rather than reported as single rows because their benchmark suites only partially overlap with ours. More importantly, they unify localization-style prediction tasks, whereas STITCH also includes retrieval and frame-budgeted VLM reasoning. Types: Sup. = supervised, Unsup. = unsupervised, ZS = zero-shot transfer, TF = training-free.
Figure 5: STITCH gains are largest when frame density is lowest. These diagnostics compare STITCH-based selection with uniform sampling for Qwen3-VL-8B [ 2 ] . Left : error by video duration, aggregated over VideoMME, MLVU, and LongVideoBench at an 8-frame budget. Center : MLVU error by frame budget. Right : MLVU error by normalized frame budget in frames/sec.
Table 2: Generic event boundary detection on Kinetics-GEBD (val) and TAPOS (val). F1 scores (%) at selected relative-distance thresholds and averaged over all ten (0.05–0.50). Sup. = supervised, Unsup. = unsupervised, ZS = zero-shot transfer (trained on a different dataset and applied without retraining), TF = training-free.
Method
Type
R@1 IoU=.3
R@1 IoU=.5
R@1 IoU=.7
mIoU
SCDM [ 42 ]
Sup.
54.80
36.75
19.86
–
2D-TAN [ 45 ]
Sup.
59.45
44.51
26.54
–
SimBase [ 3 ]
Sup.
63.98
49.35
30.48
47.07
TFVTG [ 46 ]
TF
49.34
27.02
13.39
34.10
STITCH-Max (Ours)
TF
52.25
32.85
17.87
37.40
STITCH-Mean (Ours)
TF
51.54
32.51
17.95
37.10
Appendix
Table 3: Language-based moment retrieval on ActivityNet Captions (val split, 17,505 queries). All values are percentages. R@1 IoU= τ denotes recall at rank 1 with IoU threshold τ . † GEM with p=3 ; ‡ coherence-weighted mean with τ=5 . Sup. = supervised, TF = training-free.
Table 4: Language-based moment retrieval on QVHighlights (val split). All values are percentages. † GEM with p=12 ; ‡ coherence-weighted mean with τ=5 . Sup. = supervised, TF = training-free.
Model
Method
Size
Fr.
MLVU
LongVideoBench
VideoMME
Base
+Method
Δ
Base
+Method
Δ
Base
+Method
Δ
Qwen3-VL-8B
greedy (ours)
8B
8
50.89
60.59
+9.70
52.43
54.67
+2.24
55.63
57.70
+2.07
MMR penalty (ours)
8B
8
50.89
62.43
+11.54
52.43
55.80
+3.37
55.63
58.59
+2.96
LLaVA-OV
Frame-Voyager †‡ [ 41 ]
7B
8
58.5
65.6
+7.1
—
—
—
53.3
57.5
+4.2
KFC † [ 7 ]
7B
8
58.5
66.2
+7.7
54.5
55.6
+1.1
53.3
55.4
+2.1
BOLT † [ 23 ]
7B
8
58.9
63.4
+4.5
54.2
55.6
+1.4
53.8
56.1
+2.3
Appendix
Table 5: Multiple-choice VLM accuracy with eight input frames (%). Base : uniform; +Method : with selection; Δ : gain. Splits: MLVU dev (M-Avg) [ 48 ] , LongVideoBench val [ 38 ] , VideoMME test [ 8 ] . VLMs evaluated: LLaVA-OV [ 19 ] and Qwen3-VL [ 2 ] . † As reported in the method’s original paper; other LLaVA-OV baselines re-evaluated in that setup. ‡ Training-based. Dashes: LongVideoBench missing for Frame-Voyager in that table.
Temporal video understanding studies how models track actions, events, and changes over time to answer questions, locate moments, and predict what happens next.