Probing Large Language Models for Autonomous Driving Behavior

W2W: Language-Model-Based Trajectory Prediction with Reinforcement Learning

ZenGen: Social Mind for LLMs

As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08% overall accuracy, and none of the 17 secondary dimensions reaches the 90% near-ceiling band. For internalization, we develop ZenGen, a diagnosis-driven training recipe combining supervised fine-tuning, on-policy distillation, and rubric-based reinforcement learning. Across five social-cognition benchmarks, ZenGen consistently outperforms its base models, with ZenGen-27B-Stage2 achieving the best average score and ZenGen-32B-Stage2 remaining competitive with DeepSeek-V4-Pro. For deployment-time grounding, we build Actio, a harness-controlled inference architecture that routes four typed supports into reasoning: PRISM for procedural guidance, Starling for runtime mental-state representation, SAGE for reusable experience, and gated RAG for external social and normative knowledge. Across five base models and three benchmarks, the full harness improves 14 of 15 model-benchmark pairs and is best or tied for best in 8, demonstrating the effectiveness of typed runtime support. Together, these results show that socially intelligent LLMs require coordinated advances in evaluation, parametric internalization, and deployment-time grounding.

Language-Structured Relational Q-Learning for Threat-Aware Control in Safety-Critical Driving

This work proposes Language-Structured Relational Q-Learning, instantiated through an Ego-Centric Relational Q-Network (ERQ-Net), which jointly learns inter-vehicle relevance and action values from dynamic traffic graphs, and formalises this discrepancy as a recognition-control gap.

Multi-Branch Policy Optimization for Multimodal Large Language Models

Group-based reinforcement learning methods for multimodal large language models typically rely on trajectory-level credit assignment that applies a single advantage to all tokens in a response. However, multimodal reasoning involves substantially higher perceptual uncertainty than text-only settings, where the model must repeatedly re-examine visual information to verify intermediate interpretations, and different visual groundings can lead to divergent reasoning paths, making such uniform credit assignment particularly inadequate and causing relative advantages to progressively degenerate toward zero. To address these challenges, we propose Multi-Branch Policy Optimization (MBPO), a tree-based framework that constructs reasoning trees at vision-language decision boundaries, enabling sibling branches to explore diverse visual hypotheses and assigning segment-level credit through branch-relative advantages. We further introduce a temporal replay buffer to reuse informative segments while controlling policy staleness. Experiments on several multimodal reasoning benchmarks show that MBPO outperforms representative baselines, improving both learning signal quality and optimization efficiency. The code is publicly available at https://github.com/ShuaiLyu0110/MBPO.

Multimodal Large Language Models for Autonomous Driving: A Comprehensive Survey of Perception, Reasoning, Planning, and Safety Assurance

Autonomous driving has progressed from rule-based subsystems and modular perception, prediction, and planning stacks toward unified data-driven architectures, and multimodal large language models (MLLMs) are increasingly proposed as the cognitive substrate of the next generation of highly automated road vehicles. This 2026 survey synthesises 39 primary sources selected from an initial corpus of 274 candidate records screened over 2020-2026, organises the field around a five-role pipeline taxonomy (perception, prediction, planning, control, and human-machine interaction), and compares six representative driving MLLMs (DriveGPT-4, LMDrive, Senna, DriveLM, GPT-4V-AD, and Cosmos-1) on accuracy, latency, and parameter footprint. A benchmark coverage matrix over LingoQA, BDD-X, DriveLM, nuScenes-QA, AutoHallu, and CODA-LM exposes evaluation gaps in prediction and planning. Model behaviour is translated into safety-assurance terms by mapping four MLLM failure-mode families to the functional-safety standard ISO 26262, the Safety of the Intended Functionality standard ISO 21448 (SOTIF), and the autonomous-systems safety-case standard UL 4600. A three-tier vehicle, edge, and cloud deployment topology is described together with the digital-twin and over-the-air update infrastructure that surrounds it. The strongest empirical finding is that Cosmos-1 delivers the best accuracy among models with sub-150 ms latency (76.6 percent mean reasoning accuracy at 480 ms), leaving verifiable safety certification as the single most important open problem for closed-loop deployment. The survey closes with a six-item research agenda spanning sub-100 ms real-time inference, out-of-distribution generalisation, multi-agent intent reasoning, verifiable safety certification, long-tail corner-case coverage, and closed-loop sim-to-real transfer. The article is intended as a reference for automotive system architects, safety engineers, regulators, and machine-learning researchers preparing the next generation of automated driving systems.

Analyzing and Correcting Benevolence Bias in Large Language Models

Benevolence bias is identified and measure, a small but consistent tendency for aligned LLMs to lean toward the kinder, safer, more socially approved answer on value-laden survey questions, and is easy to diagnose and straightforward to fix.