Drivebench Eval

---
name: drivebench-eval
description: Evaluates the reliability, visual grounding, and corruption resilience of vision-language models in autonomous driving. It probes whether models genuinely interpret degraded visual inputs or rely on textual priors and hallucinated reasoning when visual cues are missing or corrupted. Use when the user wants to benchmark on DriveBench, or asks about evaluating this task. Reports GPT score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2501.04003
  bibtex_key: [REDACTED]
  confidence: high
---

# drivebench-eval

> Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data, and Metric Perspectives — Shaoyuan Xie et al. (2025) (arXiv:2501.04003, 2025)

## What this evaluates

Evaluates the reliability, visual grounding, and corruption resilience of vision-language models in autonomous driving. It probes whether models genuinely interpret degraded visual inputs or rely on textual priors and hallucinated reasoning when visual cues are missing or corrupted.

## Datasets

- **DriveBench** — total 1261; splits: test (1261)

## Metrics

- `GPT score` **(primary)** — range: percent
  - An LLM-based evaluation metric that scores model responses on both answer correctness and explanation quality. Higher scores indicate more accurate and detailed reasoning.
- `Accuracy` — range: percent
  - Standard multiple-choice accuracy calculated as the ratio of correct predictions to total questions. Random baseline is approximately 33% for perception tasks (3 choices) and 25% for behavior tasks (4 choices).

## Input / output format

**Input**: Driving scene image paired with a natural language question (covering perception, prediction, planning, or explanation tasks). Text-only inputs are tested by replacing the image with a fully black image.

**Output**: Natural language text response containing the answer and, for open-ended tasks, a detailed explanation. For multiple-choice tasks, a single option selection.

## Scoring recipe

```python
def compute_metrics(predictions, golds, task_type):
    # Accuracy for MCQs
    correct = sum(1 for p, g in zip(predictions, golds) if p.strip() == g.strip())
    accuracy = correct / len(golds) * 100
    
    # GPT Score for open-ended
    gpt_scores = []
    for pred, gold in zip(predictions, golds):
        # Call LLM judge with prompt: 'Evaluate correctness and explanation quality'
        score = llm_judge(pred, gold)
        gpt_scores.append(score)
    gpt_score = sum(gpt_scores) / len(gpt_scores)
    return accuracy, gpt_score
```

## Common pitfalls

- GPT-based scoring heavily rewards detailed explanations, which can unfairly inflate scores compared to human evaluations that only check answer correctness.
- Language metrics like ROUGE-L and BLEU-4 fail to capture critical semantic information in driving contexts and can be misleadingly high for models that learn to mimic response formats rather than reason correctly.
- Models can achieve high accuracy on text-only prompts by leveraging general knowledge or text cues (e.g., camera coordinates) instead of visual grounding, masking a lack of true multimodal capability.

## Evidence (verbatim from paper)

> Considering the GPT score also takes the quality of explanations into account in addition to answer correctness, we further analyze accuracy on MCQs to isolate the potential scoring advantages due to explanations (discussed in Sec. 3.4).

## Citation

```bibtex
@misc{xie2025arevlmsready,
  title={Are VLMs Ready for Autonomous Driving? An Empirical Study from the Reliability, Data, and Metric Perspectives},
  author={Shaoyuan Xie et al. (2025)},
  year={2025},
  note={arXiv:2501.04003}
}
```

- arXiv: 2501.04003