@inproceedings{Li2026e,

title = {GeoFlex: Geographic-Prior Flexible Adaptation of Per-Domain Diffusion Trajectory Planners to Held-Out Deployment Regions},

author = {Yun Li and Ehsan Javanmardi and Simon Thompson and Jiachen Gong and Qunli Zhang and Zifan Zeng and Shiming Liu and Peng Wang and Yidu Zhang and Manabu Tsukada},

year  = {2026},

date = {2026-11-09},

urldate = {2026-11-09},

booktitle = {Conference on Robot Learning (CoRL 2026)},

address = {Austin, Texas, USA},

abstract = {Cross-region deployment of a structured multi-domain diffusion planner reduces to a small-data fine-tuning problem rather than a per-scene routing problem. We test eight zero- or low-training routing schemes on three real-world held-out maps under a disjoint-lanelet protocol; none beats a trivial static rule that picks the geographically nearest training domain. A Fano bound and a split-conformal exchangeability rejection (p=0.0035) explain why no scene-feature router can break this static-prior ceiling, and a PAC-Bayes-certified von Mises-Fisher mixture, the strongest principled routing-class baseline, confirms the limit is architectural: any method confined to the linear span of the four per-expert weights is bounded by the static-prior fleet mean (15.85 vs. 13.74 m). Stepping outside this subspace breaks the ceiling. Joint replay, mixing target-region scenes with training scenes at 10:1 during fine-tuning, reaches 4.26 m fleet-mean on disjoint lanelets (-76% vs. frozen substrate) at +0.74 m in-region regression; the recipe replicates across all four nuPlan leave-one-city-out configurations including a -35% Singapore reduction under left-hand-traffic rule shift. As a methodological contribution we introduce variance-collapse forensics: a two-line diagnostic (within-scene K-sample variance plus cross-scene pairwise L2) that unmasks hypernetwork-conditioned LayerNorm and faithful TENT (test-time entropy minimisation) as near-stationary degenerate trajectories whose seemingly-strong fleet means would otherwise overstate the gain by an order of magnitude. The same backbone has passed real-vehicle closed-loop deployment under a production DPO LoRA (joint-replay weights are not yet on-vehicle), where adversarial-aware training lifts peak deceleration 3.4x without losing natural-data comfort.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Yun2026,

title = {An Open-Source Modular Benchmark for Diffusion-Based Motion Planning in Closed-Loop Autonomous Driving},

author = {Yun Li and Simon Thompson and Yidu Zhang and Ehsan Javanmardi and Manabu Tsukada},

year  = {2026},

date = {2026-09-15},

urldate = {2026-09-15},

booktitle = {The IEEE International Conference on Intelligent Transportation Systems (ITSC2026)

},

address = {Naples, Italy},

abstract = {Diffusion-based motion planners have achieved

state-of-the-art results on benchmarks such as nuPlan, yet

their evaluation within closed-loop production autonomous

driving stacks remains largely unexplored. Existing

evaluations abstract away ROS 2 communication latency and

real-time scheduling constraints, while monolithic ONNX

deployment freezes all solver parameters at export time. We

present an open-source modular benchmark that addresses

both gaps: using ONNX GraphSurgeon, we decompose a

monolithic 18,398-node diffusion planner into three

independently executable modules and reimplement the

DPM-Solver++ denoising loop in native C++. Integrated as a

ROS 2 node within Autoware, the open-source AD stack

deployed on real vehicles worldwide, the system enables

runtime-configurable solver parameters without model

recompilation and per-step observability of the denoising

process, breaking the black box of monolithic deployment.

Unlike evaluations in standalone simulators such as CARLA,

our benchmark operates within a production-grade stack and

is validated through AWSIM closed-loop simulation. Through

systematic comparison of DPM-Solver++ (first- and

second-order) and DDIM across six step-count configurations

(N in {3, 5, 7, 10, 15, 20}), we show that encoder caching

yields a 3.2x latency reduction, and that second-order

solving reduces FDE by 41% at N=3 compared to first-order.

The complete codebase will be released as open-source,

providing a direct path from simulation benchmarks to

real-vehicle deployment. Project page: },

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Li2025d,

title = {Multi-PrefDrive: Optimizing Large Language Models for Autonomous Driving Through Multi-Preference Tuning},

author = {Yun Li and Ehsan Javanmardi and Simon Thompson and Kai Katsumata and Alex Orsholits and Manabu Tsukada},

url = {https://liyun0607.github.io/},

doi = {10.1109/IROS60139.2025.11247608},

year  = {2025},

date = {2025-10-19},

urldate = {2025-10-19},

booktitle = {2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},

address = {Hangzhou, China},

abstract = {This paper introduces Multi-PrefDrive, a framework that significantly enhances LLM-based autonomous driving through multidimensional preference tuning. Aligning LLMs with human driving preferences is crucial yet challenging, as driving scenarios involve complex decisions where multiple incorrect actions can correspond to a single correct choice. Traditional binary preference tuning fails to capture this complexity. Our approach pairs each chosen action with multiple rejected alternatives, better reflecting real-world driving decisions. By implementing the Plackett-Luce preference model, we enable nuanced ranking of actions across the spectrum of possible errors. Experiments in the CARLA simulator demonstrate that our algorithm achieves an 11.0% improvement in overall score and an 83.6% reduction in

infrastructure collisions, while showing perfect compliance with traffic signals in certain environments. Comparative analysis against DPO and its variants reveals that Multi-PrefDrive’s superior discrimination between chosen and rejected actions, which achieving a margin value of 25, and such ability has been directly translates to enhanced driving performance. We implement memory-efficient techniques including LoRA and 4-bit quantization to enable deployment on consumer-grade hardware and will open-source our training code and multi-rejected dataset to advance research in LLM-based autonomous driving systems. Project Page (https://liyun0607.github.io/)},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Li2025c,

title = {PrefDrive: Enhancing Autonomous Driving through Preference-Guided Large Language Models},

author = {Yun Li and Ehsan Javanmardi and Simon Thompson and Kai Katsumata and Alex Orsholits and Manabu Tsukada},

url = {https://github.com/LiYun0607/PrefDrive/

https://huggingface.co/liyun0607/PrefDrive

https://huggingface.co/datasets/liyun0607/PrefDrive},

doi = {10.1109/IV64158.2025.11097672},

year  = {2025},

date = {2025-06-22},

urldate = {2025-06-22},

booktitle = {36th IEEE Intelligent Vehicles Symposium (IV2025)},

address = {Cluj-Napoca, Romania},

abstract = {This paper presents PrefDrive, a novel framework that integrates driving preferences into autonomous driving models through large language models (LLMs). While recent advances in LLMs have shown promise in autonomous driving, existing approaches often struggle to align with specific driving behaviors (e.g., maintaining safe distances, smooth acceleration patterns) and operational requirements (e.g., traffic rule compliance, route adherence). We address this challenge by developing a preference learning framework that combines multimodal perception with natural language understanding. Our approach leverages Direct Preference Optimization (DPO) to fine-tune LLMs efficiently on consumer-grade hardware, making advanced autonomous driving research more accessible to the broader research community. We introduce a comprehensive dataset of 74,040 sequences, carefully annotated with driving preferences and driving decisions, which, along with our trained model checkpoints, will be made publicly available to facilitate future research. Through extensive experiments in the CARLA simulator, we demonstrate that our preference-guided approach significantly improves driving performance across multiple metrics, including distance maintenance and trajectory smoothness. Results show up to 28.1% reduction in traffic rule violations and 8.5% improvement in navigation task completion while maintaining appropriate distances from obstacles. The framework demonstrates robust performance across different urban environments, showcasing the effectiveness of preference learning in autonomous driving applications.  },

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Yun2024,

title = {Large Language Models for Human-like Autonomous Driving Decision Making: A Survey},

author = {Yun Li and Kai Katsumata and Ehsan Javanmardi and Manabu Tsukada},

doi = {10.1109/ITSC58415.2024.10919629},

year  = {2024},

date = {2024-09-24},

urldate = {2024-09-24},

booktitle = {27th IEEE International Conference on Intelligent Transportation Systems (ITSC 2024)},

address = {Edmonton, Canada},

abstract = {Large Language Models (LLMs), AI models trained on massive text corpora with remarkable language understanding and generation capabilities, are transforming the field of Autonomous Driving (AD). As AD systems evolve from rule-based and optimization-based methods to learning-based techniques like deep reinforcement learning, they are now poised to embrace a third and more advanced category: knowledge-based AD empowered by LLMs. This shift promises to bring AD closer to human-like AD. However, integrating LLMs into AD systems poses challenges in real-time inference, safety assurance, and deployment costs. This survey provides a comprehensive and critical review of recent progress in leveraging LLMs for AD, focusing on their applications in modular AD pipelines and end- to-end AD systems. We highlight key advancements, identify pressing challenges, and propose promising research directions to bridge the gap between LLMs and AD, thereby facilitating the development of more human-like AD systems. The survey first introduces LLMs’ key features and common training schemes, then delves into their applications in modular AD pipelines and end-to-end AD, respectively, followed by discussions on open challenges and future directions. Through this in-depth analysis, we aim to provide insights and inspiration for researchers and practitioners working at the intersection of AI and autonomous vehicles, ultimately contributing to safer, smarter, and more human-centric AD technologies.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}