TTA-Vid: Generalized Test-Time Adaptation for Video Reasoning

Soumya Shamarao Jahagirdar; Edson Araujo; Anna Kukleva; M. Jehanzeb Mirza; Saurabhchand Bhati; Samuel Thomas; Brian Kingsbury; Rogerio Feris

cs.CVApr 1, 2026

TTA-Vid: Generalized Test-Time Adaptation for Video Reasoning

Authors: Soumya Shamarao Jahagirdar, Edson Araujo, Anna Kukleva, M. Jehanzeb Mirza, Saurabhchand Bhati, Samuel Thomas, Brian Kingsbury, Rogerio Feris,

Organizations: University of Tübingen · Max Planck Institute for Informatics, SIC · MIT · IBM Research ·

Abstract

Recent video reasoning models have shown strong results on temporal and multimodal understanding, yet they depend on large-scale supervised data and multi-stage training pipelines, making them costly to train and difficult to adapt to new domains. In this work, we leverage the paradigm of Test-Time Reinforcement Learning on video-language data to allow for adapting a pretrained model to incoming video samples at test-time without explicit labels. The proposed test-time adaptation for video approach (TTA-Vid) combines two components that work simultaneously: (1) a test-time adaptation that performs step-by-step reasoning at inference time on multiple frame subsets. We then use a batch-aware frequency-based reward computed across different frame subsets as pseudo ground truth to update the model. It shows that the resulting model trained on a single batch or even a single sample from a dataset, is able to generalize at test-time to the whole dataset and even across datasets. Because the adaptation occurs entirely at test time, our method requires no ground-truth annotations or dedicated training splits. Additionally, we propose a multi-armed bandit strategy for adaptive frame selection that learns to prioritize informative frames, guided by the same reward formulation. Our evaluation shows that TTA-Vid yields consistent improvements across various video reasoning tasks and is able to outperform current state-of-the-art methods trained on large-scale data. This highlights the potential of test-time reinforcement learning for temporal multimodal understanding.

Figures & tables

Figure 1: TTA-Vid adapts vision-language models at test-time by sampling multiple frame subsets, enforcing majority-consistency among generated answers, and updating a frame-importance distribution via a multi-armed bandit. This allows a test-time adaptation without labels while selecting frames most relevant for reasoning.

Figure 2: Overview of TTA-Vid: Our method performs test-time adaptation of model parameters through a batch-aware reinforcement learning objective and adaptively selects the most informative frames using a multi-armed bandit approach. Both components leverage a shared reward signal computed across multiple video frame subsets, enabling the model to jointly learn what to predict and which frames to attend to.

Model	#params	#frames	VideoMMMU	MMVU	SciVideoBench	VideoMME	LongVideoBench	Avg.
Random	-	-	$14.00$	$20.00$	$10.00$	$25.00$	$20.00$	$17.80$
Proprietary Models
GPT-4o [ 22 ]	$-$	$-$	$61.22$	$75.40$	$24.90$	$71.90$	$66.70$	$60.02$
Gemini 1.5 Flash [ 45 ]	$-$	$-$	$49.78$	$58.80$	$-$	$70.30$	$61.60$	$-$
Non-Reasoning Models
LLaVA-OneVision [ 24 ]	$7\text{B}$	64	$33.89$	$49.20$	$18.80$	$58.20$	$56.30$	$43.27$

Table 1: Performance Comparison. We compare TTA-Vid with the state-of-the-art methods on two instructional video question-answering tasks. The best results are highlighted in bold . For fair comparison, we repurpose TTRV with by training it in multi-frame setup on four frames.

TTRL	MAB Rewards	Gen. Frame Sampling	VideoMMMU				MMVU
TTRL	MAB Rewards	Gen. Frame Sampling	Perception	Comprehension	Adaptation	Avg	MMVU
$\times$	$\times$	$\times$	66.33	42.33	39.33	49.33	60.80
$\checkmark$	$\checkmark$	$\times$	73.33	46.66	37.00	52.33	64.80
$\checkmark$	$\checkmark$	$\checkmark$	74.33	48.66	38.00	53.66	65.60

Table 2: Component ablation on InternVL3-8B. Models are trained on subsets of each dataset for multiple epochs and evaluated on the full test sets, with results averaged across runs. Combining all components (highlighted) achieves the best overall performance across benchmarks.

Setup	VideoMMMU				MMVU
Setup	Perception	Comprehension	Adaptation	Avg	MMVU
InternVL3-2B	$47.33$	$28.67$	$28.33$	$34.77$	50.72
Full TTA	$48.23$	$29.27$	$31.14$	$36.21$	$53.45$
Generalized TTA	48.33	31.33	32.00	37.22	54.40

Table 3: Full vs. generalized test-time adaptation. TTA-Vid is trained on multiple subsets of 32 samples and evaluated on the same batch (Per-batch Evaluation) or on unseen batches (Generalized). Results are reported for InternVL3-2B.

Train $\downarrow$ / Test $\rightarrow$	VideoMMMU	MMVU	SciVideoBench	LongVideoBench
Base InternVL3-8b	49.33	60.80	26.50	59.08
VideoMMMU	51.48	62.08	30.20	60.81
MMVU	53.66	65.60	28.20	60.95
SciVid	50.89	61.28	28.70	59.98
LVB	50.55	61.60	28.80	61.48

Table 4: Cross-dataset generalization of InternVL3-8B (32 frames, global averaging). Rows denote training dataset and columns denote evaluation dataset. Best indicates highest performance per test set. Diagonal indicates in-domain results.

Table 7

Figure 3: Qualitative comparison of frame selection strategies. Random sampling (left) versus our learned selection (right) on two VideoMMMU (Perception) examples. (1) : Musical question which requires the model to identify the sequence of notes displayed at a certain timestamp of the video. Our method selects the critical frame (correct answer I), while random sampling misses it (predicts G). (2) : Accounting question requiring localization of the value. Our method identifies the highlighted value in blue to be (0.58, option G), while random sampling fails (predicts 2.5, option E).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Table 9

# frames	$W_{clip}$	$W_{dist}$	LongVideoBench	VideoMME	SciVideoBench
4	1.0	0.0	52.35	56.48	27.70
4	0.8	0.2	52.58	56.88	27.50
4	0.5	0.5	52.80	57.11	27.80
4	0.2	0.8	53.55	57.48	27.90
4	0.0	1.0	52.05	55.51	27.80
32	1.0	0.0	60.58	65.74	29.80

Appendix

Table 10: Frame sampling using CLIP and learned distribution. In this Table, we analyze effect of using frame sampling based on the combination of CLIP score and learned distribution on InternVL3-8b as the base model. The setup used in the paper is highlighted .

Frame Sampling	MMVU	LongVideoBench	VideoMME	SciVideoBench
Uniform	65.60	61.48	65.11	29.80
1 FPS	64.96	61.11	64.00	29.40

Appendix

Table 11: Uniform sampling vs. 1 FPS. Comparison of TTA-Vid performance using two frame sampling strategies: (a) selecting 32 frames from 40 uniformly sampled frames and (b) sampling one frame per second from the video (at least and up to 32 frames). The setup used in the paper is highlighted .

Stage	Training Dataset				Test Acc. (in%)
Stage	MMVU	VideoMMMU	SciVid	Video-MME	TTA-Vid
1	✓	-	-	-	64.48 ( $+5.28\%$ )
2	✓	✓	-	-	48.83 ( $+1.39\%$ )
3	✓	✓	✓	-	25.80 ( $+9.40\%$ )
4	✓	✓	✓	✓	61.22 ( $+5.22\%$ )

Appendix

Table 12: Continual training across datasets. Each stage initializes from the previous model and is further trained with an additional dataset. We train the model on each stage only for one epoch. Colored cells indicate the dataset used for evaluation at each stage.

Figure 4: Performance per category: LongVideoBench Orange and blue dotted lines correspond to the overall performance of base model and TTA-Vid respectively. LongVideoBench dataset has multiple categories. The highest gain over the base model are observed for the following categories: Computer Science, STEM, New programs.

Figure 5: Performance based on duration of the video: LongVideoBench Orange and blue dotted lines correspond to the overall performance of base model and TTA-Vid respectively. The highest gains are observed for videos with duration of 1-3 minutes. TTA-Vid also works well with very long videos, i.e. over 30 minutes.

Panel of Figure 6 — Figure 6: Distribution evolution. Distribution progression for uniformly sampled 40 frames, over epochs for a single example with two types of initialization: uniform initialization vs CLIP initialization.

Panel of Figure 7 — Figure 7: SciVideoBench performance analysis. Comparison between the base InternVL3-8B model and TTA-Vid model with the learned frame selection distribution. We report accuracy across (a) high-level scientific categories, (b) specific scientific disciplines, and (c) question types.

Panel of Figure 8 — Figure 8: Qualitative Examples from LongVideoBench dataset.

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 322 papers

TimeThink: Reasoning with Time for Video LLMs
Jul 6, 2026Handong Li, Longteng Guo, Zikang Liu +8Temporal Video Grounding Process Reward Models
School of Artificial Intelligence, University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · OPPO AI Center, OPPO Inc.
VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization
Jun 1, 2026Junhao Cheng, Liang Hou, Tianxiong Zhong +4Test-Time Adaptation Video Diffusion Models
City University of Hong Kong · Kling Team, Kuaishou Technology
Temporal-Aware Reasoning Optimization for Video Temporal Grounding
Jun 8, 2026Minghang Zheng, Zihao Yin, Yi Yang +2Temporal Video Grounding Temporal Reasoning
Wangxuan Institute of Computer Technology, Peking University · Central Media Technology Institute, Huawei Technologies Ltd. · State Key Laboratory of General Artificial Intelligence, Peking University.

Jul 6, 2026J/K move · Enter open · S save

TimeThink: Reasoning with Time for Video LLMs

Handong Li, Longteng Guo, Zikang Liu, Dongze Hao+7

School of Artificial Intelligence, University of Chinese Academy of Sciences · Institute of Automation, Chinese Academy of Sciences · OPPO AI Center, OPPO Inc.

Temporal Video Grounding Process Reward Models Video Reasoning RL for VLMs RL for Video Reasoning

Open paper AskPDF arXiv

Video reasoning requires models to identify and verify temporally localized evidence within long video sequences. Recent Video Large Language Models (Video-LLMs) have shown promising reasoning abilities when aligned with reinforcement learning, yet existing approaches typically rely on outcome-based rewards that supervise only the final prediction. Such supervision provides limited guidance on how models should discover the relevant temporal evidence during intermediate reasoning. In this work, we propose TimeThink, a reinforcement learning framework that explicitly guides temporal evidence discovery in Video-LLMs. Our key idea is to treat temporal clue steps as the fundamental optimization primitive of video reasoning, where each reasoning step references a candidate time interval in the video. We introduce a step-wise temporal process reward that provides localized credit assignment for these clues and a joint process--outcome optimization objective that balances reasoning fidelity with task correctness. To enable scalable training, we construct TimeThink-RFT-20K, a dataset with automatically derived temporal evidence segments. Extensive experiments across video reasoning, temporal grounding, and general video understanding benchmarks show that TimeThink consistently improves both temporal localization and reasoning performance, achieving state-of-the-art results among open-source video RL models.

Abstract

Test-Time Adaptation

Explore similar work

TimeThink: Reasoning with Time for Video LLMs

VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization

Temporal-Aware Reasoning Optimization for Video Temporal Grounding