Research map / Vision-language and multimodal / Robot action and navigation

Intend, Reflect, Refine: An Adaptive Multimodal Reflection Framework for Autonomous Driving

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes IRR-Drive, a VLA driving model that drafts a textual trajectory intention, predicts future semantic BEV tokens as a reflection space, refines the plan, and adaptively chooses between direct prediction and reflection via GSPO-based RL with a concentric OBB-FDE reward.

Problem

End-to-end VLA driving models typically output a trajectory in a single pass without explicitly checking its future consequences, and existing textual or visual reasoning is either too coarse or not used as an explicit plan-verification space; reflection also adds computational cost in simple scenes

Approach

Use predicted future semantic BEV representations, conditioned on the current scene and an initial textual trajectory intention, as a structured multimodal reflection space for refining the plan; train the model to adaptively reflect only in challenging scenes via an adaptive reflection reward and a geometric concentric OBB-FDE reward.

Key results

• IRR-Drive-4B reaches 91.3 PDMS on NAVSIM v1, the highest among compared methods (ELF-VLA-8B 91.0, LaST-VLA 91.3 tied).
• IRR-Drive-4B achieves 89.0 EPDMS on NAVSIM v2, versus 87.1 for the previous best ELF-VLA-8B and LaST-VLA-8B.
• Adaptive reflection reduces runtime to 1.70s versus 3.03s for always-reflect while reaching 91.3 PDMS vs 90.8 for always-reflect.

Surprising findings

• Uniform reflection across all scenes gives only +0.2 PDMS over non-reflection while increasing runtime more than 2x.
• Training RL on 15k score-and-variance-selected samples outperforms RL on the full 132k reflection dataset (91.3 vs 87.7 PDMS).

Limitations

• Not fully suitable for real-time inference, per the authors.
• Relies on simulator-generated future BEV and open-source LLM/VLM-generated reflection labels.
• Reported PDMS gains over LaST-VLA on NAVSIM v1 are negligible (91.3 vs 91.3) and some EPDMS sub-metrics (e.g., comfort in EC) are lower than baselines.

Experimental setup

Datasets: NAVSIM v1 navtrain DriveLM NuInstruct nuScenesQA CODA-LM NAVSIM BEV reconstruction/prediction data

Benchmarks: NAVSIM v1 NAVSIM v2

Metrics: PDMS EPDMS NC DAC TTC Comfort EP runtime reflection ratio

Baselines: AutoVLA AdaThinkDrive ELF-VLA FSDrive LaST-VLA DriveVLA-W0 ReCogDrive DiffusionDrive Hydra-MDP++ DriveSuprim Senna-2 WAM-Flow

Builds on: GSPO Qwen2.5-VL Qwen3-VL NAVSIM DriveVLA-W0

Closest papers by meaning

• MindDriver: Introducing Progressive Multimodal Reasoning for Autonomous Driving (CVPR 2026 · similarity 0.90)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.90)
• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.89)
• Future-Aware End-to-End Driving: Bidirectional Modeling of Trajectory Planning and Scene Evolution (NeurIPS 2025 · similarity 0.89)
• SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving (CVPR 2026 · similarity 0.88)
• DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.88)