TVBench

Lost in Time: A New Temporal Benchmark for VideoLLMs

Daniel Cores*, Michael Dorkenwald*, Manuel Mucientes, Cees G. M. Snoek, Yuki M. Asano

*Equal contribution · BMVC 2025

2,525 questions · 10 temporal tasks · Multiple-choice video QA

Download · Evaluate your model · Results · Cite TVBench

TVBench evaluates how well video-language models understand what happens over time: event order, repetitions, object identity, movement and temporal localization. Its ten tasks use multiple-choice questions with fixed reference answers, enabling objective scoring without an LLM judge.

Updates

• 13 November 2026 — Expanded results. 47 completed video-model measurements, including direct and thinking settings: 1 original v3 result and 46 newer results. Macro and micro scores are shown separately. Paper v4 is in preparation. This update adds model results; the benchmark questions and labels are unchanged.
• 23 December 2024 — Annotation correction. Unexpected Action labels were updated. If you downloaded an earlier copy, please redownload the annotations.

What TVBench measures

TVBench was designed to reduce three common shortcuts: solving video questions from a single frame, guessing from the wording of the answer choices, and relying on general knowledge instead of evidence in the video.

Download

1. Download the pinned dataset and extract its video archives.

The commands below download the frozen snapshot, then extract each ZIP beside its archive. This includes ZIPs inside video/egocentric_sequence/. Run them in the folder where you want to create TVBench/. The Hugging Face CLI documentation describes the download options.

python -m pip install --upgrade huggingface_hub
hf download FunAILab/TVBench \
  --repo-type dataset \
  --revision 1d933b2aced8da5d6b315bb6a37807d126e05526 \
  --local-dir TVBench

python - <<'PY'
from pathlib import Path
from zipfile import ZipFile

for path in sorted(Path("TVBench/video").rglob("*.zip")):
    with ZipFile(path) as archive:
        archive.extractall(path.parent)
PY

2. Add the NTU RGB+D clips for Action Antonym.

These clips are distributed separately. Obtain the TVBench subset through the official NTU RGB+D website; downloading the full NTU dataset is unnecessary. Put the clips in TVBench/video/action_antonym/.

TVBench/
├── json/
│   ├── action_antonym.json
│   └── ...
└── video/
    ├── action_antonym/    # NTU RGB+D clips downloaded separately
    └── ...

For reproducible evaluation: use annotation revision 1d933b2aced8da5d6b315bb6a37807d126e05526. The Hub viewer calls the split train; these examples are the benchmark evaluation set.

Each task's json/ file contains the question, candidate answers, reference answer and video path. Video paths are relative to video/<task>/. The source datasets are Perception Test, CLEVRER, STAR, MoVQA, Charades-STA, NTU RGB+D, FunQA and CSV; see the source notices.

Evaluate your model

TVBench task-macro accuracy (%) is the mean of the ten task accuracies, with each task weighted equally. Report question-micro alongside it.

1. Evaluate all 2,525 questions across all ten tasks using the frozen annotation revision. Preserve the permitted video intervals. A partial run or task subset is not a completed full-benchmark result.
2. For each task, divide its number of correct answers by its full number of questions. Invalid or unparseable answers count as incorrect; do not remove them from the denominator.
3. Take the unweighted mean of those ten task accuracies, then multiply by 100. Use unrounded counts/accuracies for this calculation and round only the final reported percentage.

# correct[t] and total[t] are counts for each of the ten complete tasks.
tvbench_macro_percent = 100 * sum(
    correct[t] / total[t] for t in all_10_tasks
) / 10
tvbench_micro_percent = 100 * sum(correct.values()) / sum(total.values())

Using the original notebook? Its “Total Acc” output is question-micro accuracy. Calculate the headline task-macro score from per-task counts using the formula above.

Share your result: open an issue with:

• Overall macro and micro scores, plus correct/total counts for all ten tasks. The public table displays only the two overall scores.
• The dataset annotation revision and model/checkpoint version.
• FPS or fixed frame count, frame cap, and resolution/pixel budget.
• Thinking mode/effort, answer-extraction rule, and a paper/report or evaluation-code link. State any selection across settings or reasoning modes.

Predictions may be shared where permitted. Missing settings remain marked --. Every imported score retains its reporting-source citation.

Results

47 completed video results scoring ≥50% (author macro or published overall). Models are grouped by family; direct and thinking settings are adjacent.

Macro: equal weight per task. Micro: equal weight per question. Both author scores use the same configuration and all 2,525 questions. Original v3 and Tarsier2 macro scores are calculated from rounded task results. Dataset versions and input budgets may differ. Scoring instructions.

Source: Ours = our evaluation; Ours [v3] = the original TVBench table. Other Source citations link to the reporting paper. Model names link to the model paper/release.

Setting: Direct = native/formal thinking disabled; Thinking = enabled; CC = constrained-choice decoding; EVS50 = 50% visual-token pruning; -- = not established. Input budgets vary across models, so these scores do not represent equal compute.

Model / checkpoint	Model setting: frames/FPS · resolution · reasoning	Macro (%)	Micro (%)	Source
Gemini
Gemini 4 Argon	2 fps; native video; HIGH media (provider preprocessing); Thinking (medium)	86.82	87.56	Ours
Gemini 3.8 Flash	2 fps; native video; HIGH media (provider preprocessing); Thinking (medium)	82.59	82.30	Ours
Gemini 3.5 Flash-Lite	2 fps; native video; HIGH media (provider preprocessing); Thinking (medium)	66.81	68.20	Ours
GPT
GPT-6.1 Sol	2 fps; images; Native images; detail=original; Thinking (medium)	86.71	88.63	Ours
Qwen
Qwen3.8-27B	2 fps≤512f; Native/dynamic; Direct	67.52	67.13	Ours
Qwen3.8-Flash-Next	2 fps≤512f; Native/dynamic; Direct; CC	65.71	64.67	Ours
Qwen3.5-27B	2 fps≤512f; Native/dynamic; Direct	67.48	67.13	Ours
Qwen3.5-27B	2 fps≤512f; Native/dynamic; Thinking	74.91	74.69	Ours
Qwen3.5-397B-A17B	2 fps≤512f; Native/dynamic; Direct	69.88	70.18	Ours
Qwen3.5-122B-A10B	2 fps≤512f; Native/dynamic; Direct	69.01	68.59	Ours
Qwen3.5-35B-A3B	2 fps≤512f; Native/dynamic; Direct	68.66	68.40	Ours
Qwen3.5-4B	2 fps≤512f; Native/dynamic; Direct	65.98	66.85	Ours
Qwen3.5-9B	2 fps≤512f; Native/dynamic; Direct	64.08	63.72	Ours
Qwen3-VL-235B-A22B-Instruct	2 fps≤512f; Native/dynamic; Direct	72.51	73.23	Ours
Qwen3-VL-235B-A22B-Thinking	2 fps≤512f; Native/dynamic; Thinking	69.93	70.46	Ours
Qwen3-VL-8B	2 fps≤512f; Native/dynamic; Direct	61.86	61.98	Ours
Qwen3-VL-4B	8 FPS; ≤2048 frames; ≤448×448; ≤80K visual tokens	--	53.70	S4
InternVL
InternVL3.5-241B-A28B	16f; Native/dynamic; Direct; CC	73.52	72.87	Ours
InternVL3.5-8B	16f; 448×448/frame; Direct; CC	64.66	63.17	Ours
InternVL3.5-4B	64 uniform frames	--	57.10	S4
GLM
GLM-4.6V	dynamic≤640f; Native/dynamic; Direct	67.02	67.56	Ours
GLM-4.6V-Flash	dynamic≤640f; Native/dynamic; Direct	63.12	63.80	Ours
GLM-4.6V-Flash	dynamic≤640f; Native/dynamic; Thinking	66.96	67.17	Ours
Molmo2
Molmo2-4B	2 fps≤384f; 378×378/frame; Direct	65.83	66.89	Ours
Molmo2-8B	2 fps≤384f; 378×378/frame; Direct	64.20	65.82	Ours
Gemma4
Gemma4-31B	1 fps≤60f; Native/dynamic; Direct	54.79	54.53	Ours
Gemma4-31B	1 fps≤60f; Native/dynamic; Thinking	65.61	65.03	Ours
Gemma4-26B-A4B	32f; Native/dynamic; Direct	51.00	50.34	Ours
PLM
PLM-8B	32f; Native/dynamic; Direct	64.45	66.73	Ours
PLM-3B	32f; Native/dynamic; Direct	57.44	60.48	Ours
MiniCPM
MiniCPM-V-4.7-35B-A3B	1 fps≤128f; Native/dynamic; Direct	62.02	62.65	Ours
MiniCPM-V-4.5-8B	2 FPS; ≤540 frames; groups of 3	--	53.00	S4
LongVU
LongVU + RRPO	1 fps≤100f (paper); 384×384 + 378×378 (code)	--	56.50	S12
LongVU + DPO	1 fps≤100f (paper); 384×384 + 378×378 (code)	--	54.30	S12
LongVU	1 fps≤100f (paper); 384×384 + 378×378 (code)	--	53.70	S12
Tarsier
Tarsier2-7B	16f (code); ≤460,800 px (code)	52.61	54.70	S13, S14
Tarsier2-Recap-7B	--	50.38	54.00	S14
Nemotron
Nemotron-Nano-12B-v2-VL	2 fps≤128f; 512×512/frame; Direct; EVS50	53.71	54.30	Ours
Nemotron-3.5-Super-VL-120B-A12B-NVFP4	1 fps≤128f; Native/dynamic; Direct	50.73	52.20	Ours
LLaVA
LLaVA-Video-7B + RRPO (32f training variant)	64f; 384×384 (code)	--	52.20	S12
LLaVA-Video 72B	--	50.00	--	Ours v3
Other models
Intern-S2-397B	2 fps≤512f; Native/dynamic; Thinking	77.87	77.19	Ours
Eagle2.5-8B	2 FPS; ≤512 frames	--	62.00	S4
InternVideo2.5-8B	128–512 uniform frames; full video	--	58.10	S4
VideoChat3-4B	8 FPS; ≤2048 frames; ≤448² pixels/frame; ≤80K visual tokens	--	55.40	S4
VideoLLaMA3-7B	1 fps≤180f; Native/dynamic; Direct	53.90	54.61	Ours
Video-R1	2 fps≤32f; ≤200,704 px/frame (author-confirmed); Thinking	53.60	53.54	S16

How to read input settings

• Frames/FPS: f means sampled frames; ≤ denotes a maximum. FPS is nominal and can decrease when the frame cap binds.
• Resolution: “Native/dynamic” refers to the model-specific processor. A pixel budget is a maximum area, not a fixed width and height. Visual tokens and compute vary across models.
• Evidence: (recipe) and (code) label published recipes or defaults whose exact historical use is unconfirmed. -- means the setting was not established.
• Reasoning: Direct disables native/formal thinking; some direct instructions allow explanations. Thinking enables it. CC is constrained-choice decoding; EVS50 is 50% visual-token pruning.
• Precision: Local author checkpoints use BF16, except the explicitly named Nemotron NVFP4 checkpoint.

Original v3 input settings are being recovered from the evaluation owners; current model defaults have not been substituted. Selected v4 configurations reflect the best collected eligible setting under the manuscript selection, not a held-out or globally optimal configuration.

Score sources and protocol notes

• v3: original TVBench table, chronological video inputs only.
• S4: VideoChat3.
• S12: RRPO.
• S13: Tarsier2.
• S14: Tarsier2-Recap model release.
• S16: Visual Jigsaw Post-Training Improves MLLMs.

Citing a score source establishes where the number was reported. It does not establish identical annotations, preprocessing or aggregation. The benchmark changed from 2,654 to 2,525 questions and later revised labels without changing the count. Original v3 macro values here are recomputed from the ten rounded task scores; the original table and its published averages remain archived. Video-R1 macro comes from author-supplied task results; its micro and the displayed VideoChat3-report micro scores are author-confirmed. RRPO micro scores follow the released scorer linked by the author. Reports with unconfirmed averaging are omitted from this card. Tarsier2/Recap micro is supported by released output; their macro values are approximate means of rounded task results.

Thinking rows can also change answer instructions, sampling/decoding and output allowances. Their differences are configuration comparisons, not an isolated thinking-switch effect. Per-example visual tokens and compute are model-specific.

Our updated scores use the audited explicit-choice parser: clear choices are accepted; ambiguous or incomplete answers remain incorrect.

Citation

If you use TVBench, please cite Lost in Time: A New Temporal Benchmark for VideoLLMs (BMVC 2025).

@inproceedings{Cores_2025_BMVC,
  author = {Daniel Cores and Michael Dorkenwald and Manuel Mucientes and Cees G. M. Snoek and Yuki M. Asano},
  title = {Lost in Time: A New Temporal Benchmark for VideoLLMs},
  booktitle = {BMVC},
  year = {2025},
  url = {https://arxiv.org/abs/2410.07752}
}

Suggested methods sentence: “We report TVBench task-macro and question-micro accuracy over all 2,525 questions. Macro accuracy weights the ten tasks equally; micro accuracy weights questions equally. Input and reasoning settings are reported alongside the score.”

When quoting this results update, give its snapshot date (13 November 2026) and dataset URL. When reusing a score from another paper, cite that reporting paper too.