A poor start lowers the ceiling, but does not negate the value of finishing well. In autoregressive video temporal grounding, models systematically exhibit an endpoint performance asymmetry, predicting ends significantly worse than starts. A committed flawed start acts as an immutable physical constraint, irreversibly capping the attainable overlap as a sunk loss. While models should minimize the remaining conditional regret, standard post-training paradigms fail to teach this adaptation. Specifically, teacher-forced SFT suffers from strict prefix exposure bias, and holistic trajectory evaluation in standard GRPO inadvertently penalizes conditionally optimal ends if they follow a poor start. To address this, we introduce FinishStrong, a novel two-stage post-training framework explicitly designed to teach optimal continuations. First, Online-start SFT exposes the end prediction to the model's own generated contexts. Second, Conditional-tree GRPO restructures the sampling topology, centering end rewards exclusively within the same-start branch to mathematically isolate the conditional end quality from the inherited start penalty. FinishStrong-9B achieves a state-of-the-art 61.7% overall IoU on TimeLens-Bench, remarkably outperforming flagship closed-source models like Gemini-3.8-Flash and Qwen3.8-Max. Crucially, when forced to continue from imperfect starts, it reduces the subsequent end prediction error by 36.5%, proving its exceptional ability to end well after flawed starts. Code will be open.
Figures & tables
Figure 1: (a) Even strong models localize ends worse than starts. (b) Ending quality is sensitive to the start prefix. (c) A bad start lowers the ceiling, but we can still end well.
Figure 2: How standard paradigms interact with optimal continuations. (a) SFT typically trains on perfect starts, leaving generated errors less covered. (b) GRPO evaluates trajectories holistically. If a flawed start lowers the ceiling below the group mean ( Cmax<Rˉ ), even an optimal end ( Rege=0 ) might still receive a negative advantage. (c) Video-OPD provides feedback that is naturally influenced by the teacher’s own conditional capabilities.
Figure 3: FinishStrong : learn to end well after imperfect starts. (a) Online-start SFT teaches ends under detached model starts, retaining ground-truth supervision. (b) Conditional-tree GRPO samples 4 starts, each with 3 ends. Start advantages ( As ) use average downstream value; end advantages ( Ae ) center exclusively within their branch.
Table 1: Comparison with proprietary and open-source models on temporal video grounding benchmarks, TimeLens-Bench Zhang et al. (2026) (%). R@ τ is Recall@1 at IoU threshold τ ; ALL is sample-averaged IoU. Bold marks the best results.
Table 2: End prediction under shared start prefixes. Each direction averages six GT-span normalized offsets. Raw IoU averages interval overlap without ceiling normalization; cIoU measures the fraction of the fixed-start ceiling attained. Overlap and accuracy are percentages; MAE is in seconds. At the GT start, cIoU equals raw IoU.
Figure 4: (a) Absolute ceiling-normalized IoU (cIoU) and (b) Absolute end accuracy at tolerance 0.1L⋆ . Footer cards report the mean percentage-point gains over TimeLens for earlier, later, and all nonzero offsets.
Table 3: Training-stage ablations on Qwen3.5-9B. Free IoU denotes standard unconstrained grounding. For conditional metrics (cIoU, end accuracy, and end MAE), values are macro-averaged over the twelve nonzero GT-span normalized offsets. IoUs and accuracy are in percentages; MAE is in seconds. C represents TimeLens, while H is our complete FinishStrong .
Table 4: All-nonzero GT-span summaries. All columns use the same twelve nonzero GT-span-normalized offsets, providing a global average corresponding to the directional splits in Table 2 .
cIoU (%) ↑
End MAE (s) ↓
Model
Earlier
Later
Nonzero
Earlier
Later
Nonzero
SFT
45.41
68.43
56.92
20.80
6.36
13.58
GRPO
57.96
68.89
63.43
14.69
4.83
9.76
Video-OPD
59.70
65.88
62.79
13.89
5.22
9.55
TimeLens
67.96
69.16
68.56
14.15
5.68
9.92
FinishStrong
69.95
74.14
72.04
10.94
4.29
7.61
Appendix
Table 5: Video-span-normalized offset performance. Directional means average six offsets, and nonzero means average twelve. Positive offsets are shared with the GT-span scan.
Table 6: End localization across tolerances (%). Every column averages all twelve nonzero GT-span normalized offsets.
Figure 5: A consistent end-localization advantage. Each cell is FinishStrong minus the indicated baseline in Acc.1 , in percentage points. All six earlier-start conditions from each scan are included; the six identical later-start conditions appear once. Every cell is positive. The complete grid, rather than a selected hard-prefix subset, determines the comparison.
Figure 6: Video-span stress test. Complete cIoU and seconds-MAE curves for FinishStrong-9B and four controls. The negative coordinate sweeps the pre-event range; the positive coordinate is shared with the GT-span scan.
Figure 7: Unnormalized interval quality and end error. Full GT-span curves for raw IoU and seconds MAE, complementing the cIoU and end-accuracy curves in Figure 4 . Every nominal offset is retained.
Figure 8: Conditional results by dataset. Six earlier and six later GT-span normalized offsets are averaged within each dataset. All entries are measured for FinishStrong-9B. The figure reports absolute conditional performance for each dataset.
Table 8: Grounding with smaller baselines on TimeLens-Bench (%). Bold marks the best.
MAE (s) ↓
Method
IoU ↑
Start
End
Qwen3.5-4B
TimeLens
57.80
7.06
9.87
Video-OPD
56.96
7.22
9.56
FinishStrong
59.79
6.62
8.91
Qwen3.5-2B
Appendix
Table 9: Complete endpoint metrics for compact models. IoU is a percentage; start/end MAE is in seconds. Bold marks the best value within each backbone and metric; higher is better for IoU and lower is better for MAE.
Family
Free end MAE
GT-start end MAE
Paired reduction [95% CI]
Qwen3-VL-8B
12.83
8.17
4.65 [4.10, 5.22]
Qwen3.5-9B
11.18
7.38
3.81 [3.33, 4.31]
Appendix
Table 10: Exploratory GT-start intervention. End MAE and paired reductions are reported in seconds. The paired reduction isolates the exact subset of queries where both free and conditioned generations were validly parsed, ensuring a strict apples-to-apples comparison. Confidence intervals (95%) are derived via video-cluster resampling.
Training Recipe
Free IoU
Shared GT-Span Start
(%) ↑
Conditional End MAE (s) ↓
Standard SFT Baseline
56.5
7.32
SFT + 0.5× Artificial Noise
56.6
7.28
SFT + 1× Artificial Noise
56.2
7.42
Online-start SFT (Ours)
58.4
5.92
Appendix
Table 11: Impact of artificial noise injection during SFT. Evaluated on Qwen3.5-9B under identical training budgets. Conditional End MAE averages the twelve nonzero GT-span normalized offsets.
Method
Charades
ActivityNet
QVHighlights
ALL
SFT
55.4±0.18
54.4±0.21
63.5±0.14
56.2±0.15
GRPO
54.6±0.24
57.1±0.22
66.3±0.27
57.7±0.22
TimeLens
56.9±0.22
55.6±0.25
66.7±0.18
57.9±0.20
Video-OPD
53.3±0.31
58.2±0.29
70.1±0.34
58.4±0.26
FinishStrong
58.5±0.26
60.4±0.23
71.3±0.19
61.5±0.15
Appendix
Table 12: Statistical robustness on TimeLens-Bench. Performance is reported as mean ± standard deviation across three independent random training seeds for the Qwen3.5-9B models.
Figure 9: Behavior under an early start. All models begin erroneously at 3 s (GT: 30 – 52 s). While the baselines close the interval almost immediately ( 0% overlap), FinishStrong pinpoints the exact annotated end, achieving the optimal remaining ceiling ( 44.9% IoU).
Figure 10: Behavior under a late start. Faced with starts that heavily truncate the event (GT: 5 – 67 s), the baselines exhibit high variance in their end predictions. FinishStrong aligns precisely with the GT end, securing the maximum attainable overlap ( 59.7% IoU).
Temporal sentence grounding in videos (TSGV) aims to localize a temporal segment that semantically corresponds to a sentence query from an untrimmed video. Most current methods adopt pre-trained query-agnostic visual encoders for offline feature extraction, and the video backbones are frozen and not optimized for TSGV. This leads to a task discrepancy issue for the video backbone trained for visual classification, but utilized for TSGV. To bridge this gap, we propose a fully end-to-end paradigm that jointly optimizes the video backbone and localization head. We first conduct an empirical study validating the effectiveness of end-to-end learning over frozen baselines across different model scales. Furthermore, we introduce a Sentence Conditioned Adapter (SCADA), which leverages sentence features to train a small portion of video backbone parameters adaptively. SCADA facilitates the deployment of deeper network backbones with reduced memory and significantly enhances visual representation by modulating feature maps through precise integration of linguistic embeddings. Experiments on two benchmarks show that our method outperforms state-of-the-art approaches. The code and models will be released.
Allen He, Qi Liu, Kun Liu +2
BASIS International School Park Lane Harbour · UCAS · JD Explore Academy +1
Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional W1 between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.
Yuhan Zhu, Changlian Ma, Xiangyu Zeng +12
Nanjing University · Shanghai AI Laboratory · Shanghai Jiao Tong University +3
Multi-modal Large Language Models (MLLMs) have achieved remarkable progress in video temporal grounding with reinforcement learning for generating reasoning paths. However, existing models often produce superficial reasoning, which offers limited guidance for precise temporal localization. This limitation stems from (1) inefficient random exploration and (2) reward functions that focus solely on the answer correctness while ignoring reasoning quality. To address these issues, we propose TaRO (Temporal-Aware Reasoning Optimization), a framework that explicitly enhances the model's ability of thinking with time. First, we introduce a Constructive Reasoning Exploration that leverages pre-generated dense captions to construct reasoning paths grounded in explicit visual cues and timestamps, enabling efficient exploration of high-quality time-aware reasoning. Second, to evaluate reasoning quality, we design a Temporal-Sensitivity Reward. High-quality reasoning should be anchored to specific events and timestamps. If the event boundary under thinking is disrupted, such reasoning should become invalid, leading to a drop in the logit of the reasoning path. We utilize this drop as a critique of reasoning quality. Finally, TaRO follows a progressive curriculum, which starts by utilizing this reward to select better constructed reasoning paths, and evolves to a free exploration phase where the model autonomously generates effective reasoning. Experiments demonstrate that TaRO achieves state-of-the-art performance on VTG benchmarks. Code is available at https://github.com/oceanflowlab/TaRO.
Minghang Zheng, Zihao Yin, Yi Yang +2
Wangxuan Institute of Computer Technology, Peking University · Central Media Technology Institute, Huawei Technologies Ltd. · State Key Laboratory of General Artificial Intelligence, Peking University.