LAST-V2X: Language-Driven Adaptive Scenario Synthesis and Testing for Communication-Induced Failures in V2X Safety Applications
LAST-V2X addresses the problem of testing how delayed, missing, or corrupted V2X messages compromise safety-critical driving decisions. Its core idea is a framework that translates natural-language test descriptions into executable CARLA scenarios, then causally verifies whether any observed violations are actually due to degraded communication rather than ordinary traffic risk. This matters because existing adversarial methods perturb only the physical world, whereas V2X failures propagate from the information layer and must be isolated through paired counterfactual execution.
LAST-V2X is a methodologically sound contribution that tightly couples LLM-driven generation with causal verification and adaptive refinement. The paper demonstrates reliable scenario compilation at scale and convincingly argues that paired rollouts are necessary to avoid false attribution. However, the evaluation is constrained by simplified V2X application logic and by running large-scale adaptive search on a lightweight stub backend rather than fully integrated simulation, leaving open whether the reported refinement gains hold end-to-end.
The causal verification mechanism is the strongest element. By running a shadow ground-truth ego alongside the message-driven ego, LAST-V2X strictly attributes failures only when $V_{\mathrm{msg}}=1$ and $V_{\mathrm{gt}}=0$, which correctly prevents widespread false alarms: among 112 live scenes, a divergence-only monitor would raise 60 alarms, whereas the paired rollout identifies only one. The retrieval-augmented grounding pipeline also holds up well, binding LLM outputs to validated OpenDRIVE lanes and signal semantics rather than raw simulator code, and producing 100 valid scenes with zero spurious admissions under healthy communication.
First, the large-scale refinement study—central to the adaptive claims—was conducted using a lightweight stub backend instead of the full CARLA integration, raising questions about end-to-end validity and grounding drift under search mutations. Second, communication parameters such as path-loss exponents and building blockage are representative rather than field-measured, so the absolute fault thresholds may not translate to real deployments. Third, the framework depends entirely on GPT-4o with a fixed seed, making reproduction sensitive to a proprietary model snapshot. Finally, the safety applications use deterministic logic with fixed Age-of-Information budgets, which oversimplifies the nuanced fallback behaviors found in production V2X stacks.
The evidence supports the key claims within the simulated setting. The 100-scene benchmark yields zero admitted communication-induced failures under healthy communication, confirming that the pipeline does not manufacture violations. However, the paper does not quantitatively compare its failure-discovery rate against a non-LLM baseline—such as random or grid search over the same fault space—making it difficult to isolate the marginal value of language-driven synthesis from the underlying parameter search and fault injection. The comparison to related work is fair in distinguishing physical adversarial methods from information-layer testing, but lacks a head-to-head ablation on discovery efficiency.
Reproducibility is mixed. A project page is advertised, but the main text does not reference a public repository or dataset, and the exact prompt templates, while summarized in an appendix, are not provided in full. Determinism is partially addressed via fixed scenario seeds controlling 'vehicle equipage, communication loss, RSU selection, vehicle blueprints, and retrieval tie-breaking,' yet the heavy reliance on GPT-4o at temperature 0.2 means independent reproduction depends on access to a specific proprietary model. Furthermore, the use of a stub backend for the main refinement study means the wall-clock cost and simulator configuration for full CARLA reproduction of the 50-scene suite remain unclear.
Vehicle-to-everything (V2X) communication enables connected vehicles to make safety-critical decisions, but delayed, missing, or incorrect information can compromise driving safety. Systematically testing such failures is challenging because traffic interactions, communication conditions, and application responses must be jointly configured to produce meaningful scenarios. This paper presents LAST-V2X, a language-driven, adaptive framework for automated scenario synthesis and safety testing of V2X applications. LAST-V2X leverages large language models (LLMs) to translate natural-language test descriptions into executable scenarios grounded in roadway geometry, V2X communication semantics, and vehicle behavior. The framework integrates physical traffic simulation with spatially aware communication modeling and message-level fault injection to evaluate how information degradation affects vehicle decisions. To establish whether an observed safety failure is caused by communication rather than the underlying traffic conditions, LAST-V2X employs matched executions under degraded and ground-truth information. Furthermore, an execution-guided refinement mechanism combines parameter search and LLM self-reflection to identify ineffective test configurations, improve scenario effectiveness, and produce minimal reproducible failure cases. Experiments across diverse CARLA environments and V2X safety applications demonstrate the framework's ability to generate reliable scenarios, distinguish communication-induced failures from unrelated safety violations, and substantially improve failure discovery through adaptive refinement. These results establish LAST-V2X as a systematic approach to language-driven, causally verified, and adaptive safety testing for connected vehicle systems.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.