4arXiv cs.AI (Artificial Intelligence)·
AUTOPILOT-VQA: Incident-centric dashcam benchmark for evaluating VLMs on safety-critical driving scenarios
Researchers introduce AUTOPILOT-VQA, a visual question answering benchmark focused on real-world driving incidents and near-incidents captured by dashcams. The benchmark covers safety-relevant categories including weather, road conditions, accident occurrence, impact location, and avoidability reasoning, pushing models beyond object recognition toward temporally grounded safety-aware reasoning. It is released as part of the AUTOPILOT competition at CVPR 2026 and targets evaluation of vision-language models in autonomous driving contexts.
Related guides (2)
Related events (8)
Benchmark for view-level visual evidence identification in multi-view MLLMs for autonomous driving
A new arXiv preprint introduces a multi-view visual question answering benchmark targeting evidence-source identification in autonomous driving scenarios. Given six synchronized NuScenes camera views and a question, models must identify which camera view supports the answer — not just produce a correct answer. The 122-pair benchmark spans causality, counterfactual reasoning, and intent prediction, and exposes grounding failures that answer-only evaluation misses. The work addresses a meaningful gap between answer accuracy and correct visual grounding in safety-critical multimodal systems.
Act2Answer: Benchmarking commonsense and world knowledge retention in Vision-Language-Action models
Researchers introduce Act2Answer, a protocol for evaluating how much commonsense and factual knowledge VLA models retain after fine-tuning on robotics data. The approach converts knowledge benchmark questions into tabletop object-placement episodes, yielding action-grounded success rates that reduce confounds from low-level control failures. A large-scale study of 7 VLA models and 9 VLM baselines finds that VLAs retain solid performance on simple concepts but show larger gaps on richer semantic categories compared to their source VLMs, and that VQA co-training is associated with better knowledge retention.
COMPACT-VA: Planning-aligned token compression for long-context autonomous driving
Researchers introduce COMPACT-VA, a working memory framework using conditional VQ-VAE to compress extended temporal context in vision-action autonomous driving models. Compression is conditioned on historical trajectory and a learned planning intent derived from future trajectories during training, enabling end-to-end optimization without backbone modifications. On high-signal dynamic scenarios, the method achieves 68.3% success rate (>6% improvement) with 3.3x speedup and 2.7x memory reduction over uncompressed processing.
Lost in Fog: Sensor Perturbations Expose Reasoning Fragility in Driving VLAs
This paper presents a controlled robustness study of Vision-Language-Action (VLA) models in autonomous driving, evaluating Alpamayo R1 (10B parameters) across ~18,000 inference trials under eight sensor perturbation types including noise, lighting extremes, and fog. The key finding is that Chain-of-Causation (CoC) reasoning consistency is a high-fidelity proxy for trajectory reliability: when CoC explanations change post-perturbation, trajectory deviation spikes 5.3× (r=0.99 across attack types). Enabling CoC generation is associated with 11.8% average improvement in trajectory accuracy, and degradation under noise is approximately linear (R²=0.957), while standard preprocessing defenses offer only marginal benefit.
VLESA: Vision-Language Embodied Safety Agent for Real-Time Human Activity Monitoring
Researchers introduce VLESA, a framework that monitors human activities from egocentric video and triggers real-time safety interventions when dangerous actions are predicted. The system addresses intent-dependent safety — where identical actions can be safe or dangerous depending on context — using a goal-conditioned safety Q-filter trained via GRPO and an intent-action prediction agent. On the ASIMOV-2.0 benchmark, VLESA achieves higher intervention accuracy than baselines, with the Q-filter improving action safety by over 41 percentage points through goal-conditioned constrained decoding.
WikiVQABench: A Knowledge-Grounded Visual Question Answering Benchmark from Wikipedia and Wikidata
WikiVQABench is a new human-curated VQA benchmark that requires external knowledge beyond visual perception, constructed by combining Wikipedia images, captions, and Wikidata structured knowledge with LLM-generated question candidates reviewed by human annotators. The benchmark evaluates knowledge-intensive reasoning in vision-language models, covering 15 VLMs ranging from 256M to 90B parameters. Accuracy spans 24.7% to 75.6%, indicating meaningful discrimination across model scales. The dataset and code are publicly released.
CARLA-GS: Modular pipeline for photorealistic autonomous driving corner-case synthesis using LLMs and Gaussian splatting
CARLA-GS is a new modular framework for synthesizing safety-critical corner cases for autonomous driving evaluation, decoupling visual representation (editable Gaussian scenes), semantic reasoning (multi-agent LLM for waypoint generation), and physics simulation (CARLA with PID control). The pipeline reconstructs real driving scenes from the Waymo Open Dataset, uses LLMs to identify risky interactions and generate intent-level trajectories, then re-projects simulated states into Gaussian scenes for photorealistic ego-centric rendering. The approach addresses a key gap in AV safety evaluation: generating spatiotemporally consistent, physically feasible, and semantically meaningful rare scenarios at scale.
PaSBench-Video: A Streaming Video Benchmark for Proactive Safety Warning in MLLMs
PaSBench-Video is a 740-video benchmark designed to evaluate whether multimodal large language models can issue timely, accurate safety warnings during the window between a visible danger sign and an accident. Videos span four domains (driving, healthcare, daily life, industrial production) and are annotated with frame-level risk onset and accident boundaries, requiring causal temporal reasoning rather than static scene classification. Testing 13 MLLMs reveals no model exceeds 20% on the strictest metric, with recall strongly coupled to false-positive rate (Pearson r=0.64), indicating models rely on scene-level activity cues rather than genuine hazard reasoning. Performance varies sharply by domain, with driving being particularly problematic due to visual similarity between routine and hazardous scenes.