--- license: cc-by-nc-4.0 language: - en pretty_name: DashBench size_categories: - 1K

Gabriel Manalu, Minh Son Hoang, Hyeontaek Hwang, Jaesung Lee, Hyojin Choi, Tran Tuan Phong, Seon Kwon Kim, Saad Wazir, Daeyoung Kim

School of Computing, KAIST, Daejeon, Korea

BMVC 2026

[![Paper](https://img.shields.io/badge/Paper-arXiv_2611.16767-b31b1b.svg)](https://arxiv.org/abs/2611.16767) [![Project Page](https://img.shields.io/badge/Project_Page-cornercase--dashbench.github.io-blue.svg)](https://cornercase-dashbench.github.io) [![GitHub](https://img.shields.io/badge/GitHub-Code-black.svg)](https://github.com/rielmvp/cornercase-dashbench) --- ## Abstract *Natural language explanations promised by modern Video Large Language Models (Video LLMs) offer autonomous driving systems a path toward detailed scene description and interpretable decision making on the road. While recent benchmarks have begun assessing these models on driving scenes, they mainly target routine traffic conditions and general comprehension tasks based on existing driving datasets and therefore fail to evaluate model performance on the rarest and most safety-critical driving events, also known as "corner cases". To provide a comprehensive evaluation of Video LLMs on precisely these scenarios, we introduce DashBench, a benchmark testing 25 Video LLMs on 1,709 carefully curated driving videos spanning not only collision-related corner cases but also rare non-collision ones. Through extensive prompts and open-ended question answering, we expose a critical capability gap: all evaluated models perform substantially below the threshold required for safety-critical driving applications, frequently misjudging or missing dangerous events on the road ahead and hallucinating non-existent agents or actions in their explanations.* ## Dataset Overview | | | |---|---| | Clips | 1,709 ego-view dashcam clips | | Categories | 88 (46 collision, 42 non-collision) | | Collision / non-collision | 890 / 819 clips | | Sources | MM-AU (1,395 clips), RoadSocial (314 clips) | | Frames per clip | 16, uniformly sampled (`full`) or within the corner-case window (`ccw`, 823 clips) | | Prompts per clip | 5 open-ended prompts (Q1–Q5) | Each clip was manually verified to contain a visually identifiable, self-contained corner case, annotated with keyframes (abnormal start, most observable, abnormal end), and given a one-sentence ground-truth description drafted with OpenAI o3 and refined to consensus by three annotators. ## Configurations and splits | Config | Split | Rows | Description | |---|---|---|---| | `default` | `full` | 1,709 | All clips. Frames: 16 sampled uniformly across the full clip. | | `default` | `ccw` | 823 | Corner-case-window subset: the same kind of clip, but the 16 frames are sampled inside the annotated corner-case window only (the temporal-localization experiment in the paper). | | `prompts` | `train` | 5 | The Q1–Q5 prompts, the system prompt, and the judge prompt. | One row is one clip. Every clip is evaluated with **all five prompts**, so a clip is stored once instead of once per clip–prompt pair (the benchmark has 1,709 × 5 = 8,545 clip–prompt pairs). The viewer shows the text fields. The 16 frames of each clip (original JPEGs, no re-encoding) are stored separately in `frames/*.parquet`, with the frame numbers and the annotated keyframes (abnormal start, most observable, abnormal end). Join them to the text rows by `id`. ``` frames/full-*.parquet 1,709 clips frames/ccw-*.parquet 823 clips columns: id, frames (16 images), frame_numbers, abnormal_start_frame, accident_frame, abnormal_end_frame ``` ### Fields | Field | Type | Description | |---|---|---| | `id` | string | Clip id, e.g. `MMAU_000848`, `RoadSocial_RS_1`; also the key in `frames/*.parquet` | | `source` | string | `MM-AU` or `RoadSocial` | | `source_id` | string | Id in the source dataset | | `category` | string | One of the corner-case categories | | `collision` | bool | Collision vs. non-collision corner case | | `ground_truth` | string | Human-verified description of the corner case | ## Usage ```python from datasets import load_dataset repo = "gabrieldavin23/cornercase-dashbench" meta = load_dataset(repo, split="full") # or split="ccw" frames = load_dataset("parquet", data_files={"full": f"hf://datasets/{repo}/frames/full-*.parquet"}, split="full") prompts = load_dataset(repo, "prompts", split="train") clip = meta[0] print(clip["category"], "|", clip["ground_truth"]) imgs = frames[0]["frames"] # 16 PIL images, temporal order (row i of `frames` matches row i of `meta`) for p in prompts: question = p["prompt"] # ask your Video LLM this with the 16 frames ``` ## Prompts | | Prompt | |---|---| | Q1 | What is unusual in this sequence of frames? | | Q2 | What is dangerous in this sequence of frames? | | Q3 | Describe this sequence of frames in detail. | | Q4 | JSON answer with free-text `reasoning` (**Q4R**) and a boolean `unusual_or_dangerous_event_detected` (**Q4TF**). | | Q5 | Defines a corner case with examples, then asks the model to identify and describe one. | ## Evaluation Responses to Q1, Q2, Q3, Q4R and Q5 are scored by **DeepSeekV3-as-a-judge**, which answers Yes/No to whether the response captures the same corner case as the ground truth (cause, agents responsible, and timing). Accuracy is the share of Yes verdicts. DeepSeekV3 agrees with human judgments at Cohen's κ = 0.96. Q4TF is reported separately as binary detection accuracy. The judge prompt is in the `prompts` config. ## License and source data DashBench is built from clips of [MM-AU](https://github.com/jeffreychou777/LOTVS-MM-AU) and [RoadSocial](https://roadsocial.github.io/). Use of the frames is subject to the terms of those datasets. The annotations, ground-truth descriptions and prompts are released under CC BY-NC 4.0 for non-commercial research use. ## Citation ```bibtex @misc{manalu2026evaluatingvideollmsunderstanding, title={Evaluating Video LLMs' Understanding of Corner Cases in Autonomous Driving}, author={Gabriel Manalu and Minh Son Hoang and Hyeontaek Hwang and Jaesung Lee and Hyojin Choi and Tran Tuan Phong and Seon Kwon Kim and Saad Wazir and Daeyoung Kim}, year={2026}, eprint={2611.16767}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2611.16767}, } ```