Do VLMs Really Look at the Road? A Diagnostic Evaluation for Autonomous Driving
Organizations: International Institute of Information Technology, Hyderabad, India
Abstract
Vision-language models (VLMs) are increasingly being evaluated for autonomous driving through question-answering benchmarks, yet high benchmark accuracy does not necessarily reflect visually grounded reasoning. We present a diagnostic evaluation framework that systematically probes whether VLMs base safety-critical driving decisions on visual evidence or instead rely on language priors, answer biases, and dataset regularities. Our benchmark comprises two task categories, Adverse Weather and Junctions & Intersections, with a nuScenes spatial reasoning extension, and evaluates open-source (general, driving, and spatial-reasoning) VLMs along four diagnostic protocols: image versus no-image evaluation, counterfactual weather sensitivity, linguistic robustness under paraphrasing and negation, and spatial reasoning. All ground-truth labels are derived programmatically from BDD100K and DriveLM annotations, enabling reproducible evaluation without manual labeling. Across all models, we find that competitive overall accuracy conceals systematic failure modes, including strong affirmation bias, limited sensitivity to counterfactual scene changes, near-chance performance on negated questions despite high paraphrase consistency, and confident driving decisions even without visual input. These results demonstrate that current open-source VLMs remain insufficiently grounded for safety-critical autonomous driving and underscore the need for diagnostic evaluation beyond aggregate accuracy.
Figures & tables
| POPE [ 8 ] | NegVQA [ 18 ] | NegBench [ 1 ] | DriveLM [ 14 ] | DriveBench [ 16 ] | Ours | |
| Driving | – | – | – | ✓ | ✓ | ✓ |
| No-image | – | – | – | – | ✓ † | ✓ |
| Counterfactual | – | – | – | – | ✓ ‡ | ✓ |
| Paraphrase | – | – | – | – | – | ✓ |
| Negation | – | ✓ | ✓ | – | – | ✓ |
| Spatial QA | – | – | – | ✓ | – | ✓ |
| Language-prior probing (%) | Adverse weather type (%) | Semantic question types (%) | ||||||||||||
| Model | Overall | GT=yes | GT=no | A. Gap | Rainy | Foggy | Snowy | Night | Clear | Perc. | Action | Hazard | Scene | Adv. |
| Moondream 2B [ 15 ] | 60.1 | 76.0 | 46.5 | 29.5 | 59.6 | 66.1 | 67.4 | 80.2 | 36.6 | 61.9 | 62.3 | 43.0 | 72.1 | 62.2 |
| PaliGemma 3B [ 3 ] | 66.7 | 67.8 | 65.8 | 2.0 | 61.1 | 67.1 | 65.2 | 75.5 | 67.6 | 65.2 | 73.7 | 66.3 | 70.1 | 59.1 |
| SmolVLM 2.2B [ 10 ] | 66.7 | 43.8 | 85.8 | 42.0 | 56.1 | 70.5 | 63.6 | 79.6 | 69.4 | 61.1 | 78.4 | 59.2 | 68.7 | 72.7 |
| LLaVA-OV 8B [ 7 ] | 78.3 | 74.5 | 81.5 | 7.0 | 70.6 | 76.6 | 77.4 | 93.1 | 78.7 | 72.3 | 77.9 | 74.5 | 92.8 | 79.7 |
| InternVL3 8B [ 19 ] | 75.8 | 85.8 | 67.4 | 18.4 | 73.4 | 66.2 | 72.0 | 85.6 | 82.5 | 70.6 | 79.7 | 72.8 | 85.3 | 75.3 |
| Counterfactual SCS (%) | Linguistic Variation | |||||||
| Model | Overall | Fog | Rain | Snow | Observation | Para. (%) | Neg. (%) | Observation |
| Moondream 2B [ 15 ] | 13.4 | 16.7 | 16.0 | 11.6 | Lowest SCS; answers largely weather-invariant | 80.0 | 36.6 | Stable under paraphrase; worst negation |
| PaliGemma 3B [ 3 ] | 39.3 | 39.6 | 43.1 | 36.3 | Most sensitive; strongest response to rain | 80.9 | 48.6 | High stability; near-chance negation |
| SmolVLM 2.2B [ 10 ] | 28.1 | 32.2 | 29.0 | 26.6 | Moderate; fog induces largest shift | 78.5 | 43.2 | Stable paraphrase; below-chance negation |
| LLaVA-OV 8B [ 7 ] | 32.6 | 37.3 | 36.3 | 32.0 | Strong; consistently reacts across all weather | 80.0 | 44.2 | Robust to rewording; weak polarity handling |
| InternVL3 8B [ 19 ] | 27.9 | 30.9 | 31.5 | 28.8 | Moderate; balanced across weather types | 87.1 | 52.2 | Best paraphrase among 5B; only model above chance |
| Language-prior probing (%) | Semantic question types (%) | ||||||||
| Model | Overall | GT=yes | GT=no | A. Gap | Perc. | Action | Hazard | Scene | Adv. |
| Moondream 2B [ 15 ] | 63.7 | 85.0 | 45.5 | 39.5 | 82.1 | 51.9 | 34.0 | 77.1 | 37.9 |
| PaliGemma 3B [ 3 ] | 71.1 | 73.0 | 69.6 | 3.4 | 78.3 | 61.6 | 38.1 | 82.0 | 69.7 |
| SmolVLM 2.2B [ 10 ] | 73.5 | 71.7 | 74.9 | 3.2 | 85.5 | 67.1 | 47.7 | 77.7 | 66.3 |
| LLaVA-OV 8B [ 7 ] | 73.5 | 75.4 | 72.0 | 3.4 | 87.0 | 57.1 | 52.7 | 88.9 | 56.7 |
| InternVL3 8B [ 19 ] | 80.2 | 78.8 | 81.3 | 2.5 | 86.7 | 67.3 | 80.9 | 86.7 | 72.8 |
| nuScenes Spatial Reasoning (%) | Linguistic Variation | |||||||||
| Model | Overall | Spatial | Pred. | Plan. | GT=yes | GT=no | Observation | Para. (%) | Neg. (%) | Observation |
| Moondream 2B [ 15 ] | 36.2 | 48.2 | 16.6 | 51.1 | 92.3 | 14.2 | Near-complete over-affirmation; high unclear rate | 61.1 | 54.0 | Moderate paraphrase; above-chance negation |
| PaliGemma 3B [ 3 ] | 69.3 | 64.1 | 75.5 | 61.6 | 51.9 | 75.1 | Balanced; strong on prediction | 57.6 | 49.4 | Low paraphrase stability; near-chance negation |
| SmolVLM 2.2B [ 10 ] | 72.1 | 66.4 | 85.8 | 54.3 | 41.3 | 83.5 | Extreme negation bias; defaults to no | 56.2 | 50.3 | Low paraphrase stability; near-chance negation |
| LLaVA-OV 8B [ 7 ] | 71.8 | 70.8 | 76.3 | 60.7 | 62.4 | 74.9 | Balanced; consistent across q-types | 57.8 | 40.6 | Moderate stability; worst negation |
| InternVL3 8B [ 19 ] | 72.1 | 69.6 | 76.3 | 62.3 | 67.1 | 73.2 | Balanced; ties for second-highest overall | 57.9 | 54.0 | Best paraphrase; above-chance negation |
VLMs for Autonomous Driving
Vision-language models for autonomous driving connect camera and sensor data with language to interpret scenes, predict hazards, and plan actions.
See all 221 papers