README.md

---
license: cc-by-4.0
task_categories:
  - video-text-to-text
language:
  - en
size_categories:
  - 1K<n<10K
pretty_name: ExtremeWhenBench
tags:
  - video
  - temporal-grounding
  - long-video
  - benchmark
---

# ExtremeWhenBench

Hour-scale natural-language temporal grounding benchmark.

<p align="center">
  <img src="./assets/teaser.png" alt="Short-clip vs hour-long temporal grounding" width="640">
</p>

**2,273 open-form natural-language questions** over **194 hour-long
videos** (mean 75.7 min, max 9 hr) sourced from LVBench, MLVU, and
VideoMME. The median GT event is 9 s — matched to Charades-STA's 7.1 s
— so the same event grain now sits inside a search space ~**153× larger**.

Companion to *Natural-Language Temporal Grounding in Hour-Long Videos is a
Search Problem: A Benchmark and Empirical Decomposition* —
[arXiv:2606.12300](https://arxiv.org/abs/2606.12300).
Code repository: **<https://github.com/naver-ai/ExtremeWhenBench>**.

## Status

**Annotations available now.** The reference evaluation script and the
lmms-eval task plugin will be released at the GitHub repository linked
above together with the camera-ready version of the paper.

## Introduction

At hour-scale, the binding constraint is **search, not recognition**.
Video-LLMs are bottlenecked not by localizing a nearby event, but —
given a natural-language query — by finding the right region of a long
video. Short-video benchmarks can't distinguish these failure modes;
hour-scale grounding makes the distinction central and lets the task
decompose into a **search** stage and a **localize** stage —
structurally the *retrieve-then-read* split that reshaped open-domain QA.

## Results

Best per-model mIoU on Charades-STA vs. ExtremeWhenBench (same task).
Open Video-LLMs collapse 5–120×, and a frame-level CLIP retriever
overtakes every open Video-LLM under our compute budget:

| Model                          | Charades-STA | ExtremeWhenBench | Ratio |
| ------------------------------ | -----------: | ---------------: | ----: |
| **Qwen3.5-9B**                 |    **0.579** |        **0.110** |   5.3× |
| InternVL3.5-8B                 |        0.359 |            0.003 |    120× |
| LLaVA-OneVision-7B             |        0.226 |            0.003 |    75× |
| LLaVA-NextVideo-7B             |        0.089 |            0.001 |    89× |
| GPT-5.4 (64f)                  |        0.299 |            0.013 |    23× |
| Gemini-2.5-flash (1k f)        |        0.308 |            0.053 |   5.8× |
| Gemini-3.5-flash (auto-fps)    |        0.466 |            0.115 |   4.1× |
| **CLIP ViT-L/14-336** (retrieval) |    0.332 |        **0.269** |   1.2× |

CLIP outperforms every open Video-LLM on ours while sitting in the
middle of the pack on Charades-STA — the order flips when search
dominates. A failure taxonomy attributes **85%** of Video-LLM errors to
search; a retrieve-then-ground hybrid recovers **6.7×** over the
monolithic Video-LLM.

Scaling frames alone doesn't close the gap — Qwen3.5-9B keeps climbing
out to N=2,048 frames but remains far below the retrieval baseline:

<p align="center">
  <img src="./assets/figure3_sweep.png" alt="Frame-count sweep, Qwen3.5-9B" width="560">
</p>

Queries are **open-form natural language**, not template-bound:
MATTR 0.78 vs. 0.60 (TVBench) and 0.54 (Charades-STA); 1,578 unique
4-gram stems (~25× per-question gap over TVBench).

## Loading

```python
from datasets import load_dataset
ds = load_dataset("min1321/extreme-when-bench", split="test")
print(ds[0])
```

## Schema

| Column             | Type           | Description                                    |
| ------------------ | -------------- | ---------------------------------------------- |
| `qid`              | string         | Stable question id                             |
| `video_id`         | string         | Source-corpus video id                         |
| `source_corpus`    | string         | `LVBench` / `MLVU` / `VideoMME`                |
| `question`         | string         | The grounding question                         |
| `correct_interval` | list\[float, float] | Ground-truth `[start_s, end_s]`           |
| `duration_s`       | float          | Length of the GT interval (s)                  |
| `video_duration_s` | float          | Length of the full source video (s)            |
| `event_summary`    | string         | Short paraphrase of the event                  |
| `category`         | string         | Coarse event category                          |
| `youtube_url`      | string         | YouTube URL (empty for MLVU)                   |

Predictions are `[start, end]` intervals in seconds. Metrics: **mIoU**,
**R@0.3 / R@0.5 / R@0.7**, **parse-failure rate** (parse failures count
as IoU = 0).

## Source videos

This dataset ships annotations only — download the videos from their
source corpora:

| Corpus    | Videos | Where to get it                                   |
| --------- | -----: | ------------------------------------------------- |
| VideoMME  |     89 | <https://video-mme.github.io/>                    |
| LVBench   |     67 | <https://lvbench.github.io/>                      |
| MLVU      |     38 | <https://github.com/JUNJIE99/MLVU>                |

The `source_corpus`, `youtube_url`, and `video_duration_s` columns give
you the per-video lookup.

## Evaluation

Two evaluation paths will be released at the
[companion GitHub repo](https://github.com/naver-ai/ExtremeWhenBench)
together with the camera-ready (see **Status** above):

- **Reference vLLM script** — async OpenAI client that streams video URLs
  to a vLLM serve and parses `[start, end]` predictions. Matches the
  paper's Table 4 (Qwen3.5-9B, `num_frames=768`, `enable_thinking=False`):
  mIoU 0.0469.
- **lmms-eval task plugin** — drop-in for
  [EvolvingLMMs-Lab/lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval),
  with a small adapter patch (`+49` lines) that lets the `openai` adapter
  send video as a URL so vLLM can decode server-side (mIoU 0.0485 on the
  same setup). We plan to upstream the patch + task as a PR to
  `EvolvingLMMs-Lab/lmms-eval`.

Reported numbers we will reproduce:

| Path                                       | mIoU   |
| ------------------------------------------ | ------ |
| Paper Table 4 (reported)                   | 0.053  |
| Reference vLLM script                      | 0.0469 |
| lmms-eval task                             | 0.0485 |

## Citation

```bibtex
@article{seo2026extremewhenbench,
  title   = {Natural-Language Temporal Grounding in Hour-Long Videos
             is a Search Problem: A Benchmark and Empirical Decomposition},
  author  = {Seo, Sukmin and Kim, Geewook},
  journal = {arXiv preprint arXiv:2606.12300},
  year    = {2026}
}
```

## License

Annotations: CC-BY-4.0. Copyright (c) 2026-present NAVER Cloud Corp.
Source videos are governed by the licenses of LVBench, MLVU, and VideoMME
respectively; this dataset does not redistribute them.