DRIV-EX: Counterfactual Explanations for Driving LLMs

2026/02/28 by Amaia Cardiel, Zablocki, Eloi, Eloi Zablocki +2 · 1 voice
Computer Science · #Explainable Artificial Intelligence (XAI) #Multimodal Machine Learning Applications #Topic Modeling #cs.CL

paper · pdf · doi:10.48550/arxiv.2603.00696

openalex publication_date 2026/02/28 · arxiv published 2026/02/28 · openalex created_date 2026/03/04 · arxiv updated 2026/04/21 · openalex updated_date 2026/07/28

Abstract

Large language models (LLMs) are increasingly used as reasoning engines in autonomous driving, yet their decision-making remains opaque. We propose to study their decision process through counterfactual explanations, which identify the minimal semantic changes to a scene description required to alter a driving plan. We introduce DRIV-EX, a method that leverages gradient-based optimization on continuous embeddings to identify the input shifts required to flip the model's decision. Crucially, to avoid the incoherent text typical of unconstrained continuous optimization, DRIV-EX uses these optimized embeddings solely as a semantic guide: they are used to bias a controlled decoding process that re-generates the original scene description. This approach effectively steers the generation toward the counterfactual target while guaranteeing the linguistic fluency, domain validity, and proximity to the original input, essential for interpretability. Evaluated using the LC-LLM planner on a textual transcription of the highD dataset, DRIV-EX generates valid, fluent counterfactuals more reliably than existing baselines. It successfully exposes latent biases and provides concrete insights to improve the robustness of LLM-based driving agents. The code is available at "https://github.com/Amaia-CARDIEL/DRIV_EX" .

Citations

• Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail
• Poutine: Vision-Language-Trajectory Pre-Training and Reinforcement Learning Post-Training Enable Robust End-to-End Autonomous Driving
• SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment
• Can LLMs Explain Themselves Counterfactually?
• Controlled LLM Decoding via Discrete Auto-regressive Biasing
• EMMA: End-to-End Multimodal Model for Autonomous Driving
• The Llama 3 Herd of Models
• CarLLaVA: Vision language models for camera-only closed-loop driving
• SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction
• OmniDrive: A Holistic Vision-Language Dataset for Autonomous Driving with Counterfactual Reasoning
• LLMs for Generating and Evaluating Counterfactuals: A Comprehensive Study
• Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks
• COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability
• Fight Back Against Jailbreaking via Prompt Adversarial Tuning
• Bypassing the Safety Training of Open-Source LLMs with Priming Attacks
• GPT-Driver: Learning to Drive with GPT
• MotionLM: Multi-Agent Motion Forecasting as Language Modeling
• Jailbroken: How Does LLM Safety Training Fail?
• CREST: A Joint Framework for Rationalization and Counterfactual Text Generation
• Sequential Integrated Gradients: a simple but effective method for explaining language models
• BOLT: Fast Energy-based Controlled Text Generation with Tunable Biases
• Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
• Towards Automated Circuit Discovery for Mechanistic Interpretability
• TIGTEC : Token Importance Guided TExt Counterfactuals
• Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery
• Gradient-Based Constrained Sampling from Language Models
• COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamics
• Text Counterfactuals via Latent Optimization and Shapley-Guided Search
• PICARD: Parsing Incrementally for Constrained Auto-Regressive Decoding\n from Language Models
• The Power of Scale for Parameter-Efficient Prompt Tuning
• Exploring the Role of BERT Token Representations to Explain Sentence\n Probing Results
• Polyjuice: Generating Counterfactuals for Explaining, Evaluating, and Improving Models
• Explaining NLP Models via Minimal Contrastive Editing (MiCE)
• Perturbed Masking: Parameter-free Probing for Analyzing and Interpreting BERT
• Universal Adversarial Triggers for Attacking and Analyzing NLP
• BERTScore: Evaluating Text Generation with BERT
• nuScenes: A multimodal dataset for autonomous driving
• The highD Dataset: A Drone Dataset of Naturalistic Vehicle Trajectories\n on German Highways for Validation of Highly Automated Driving Systems
• Visualizing and Understanding Neural Models in NLP
• Adam: A Method for Stochastic Optimization
• Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation
• Prefix-Tuning: Optimizing Continuous Prompts for Generation
• Prefix-Tuning: Optimizing Continuous Prompts for Generation

Discussions

• 2/ DRIV-EX: Counterfactual Explanations for Driving LLMs Counterfactual explanations: the minimal semantic edit that flips an LLM driving planner's decision. They can expose unexpected biases 📄 arxiv [bsky, 2 points, 0 comments]