Research map / Vision-language and multimodal / Multimodal LLMs

NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 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 NoRD, a reasoning-free, data-efficient VLA for driving that uses SFT on a small driving set followed by Dr. GRPO post-training to fix difficulty bias that stalls standard GRPO.

Problem

VLA-based autonomous driving relies on large datasets with reasoning (CoT) annotations and RL post-training, which is costly in data, annotation, and inference tokens. Standard GRPO yields negligible gains when applied to weak SFT policies trained on small reasoning-free datasets.

Approach

Identifies that GRPO's std normalization causes difficulty bias that disproportionately penalizes high-variance (intermediate-mean reward) driving scenarios for a weak SFT policy, and replaces GRPO with Dr. GRPO (no std normalization) plus DAPO-style asymmetric clipping and no KL term, with trajectories tokenized via k-disc clustering into 2048 tokens.

Key results

• On NAVSIM navtest, NoRD-base with GRPO reaches 77.18 PDMS (+0.67%) while Dr. GRPO reaches 85.62 (+11.68%) from the 76.66 SFT baseline.
• NoRD achieves 85.6 PDMS on NAVSIM with 3 RGB frames and no reasoning, and NoRD-BoN reaches 92.4, above AutoVLA-BoN's 92.1.
• On WaymoE2E test, NoRD attains RFS 7.709 without reasoning or ensembling, ranking third among VLAs, with ADE@3 of 1.2504.

Surprising findings

• Standard GRPO yields only +0.67% on a weak reasoning-free SFT policy, while Dr. GRPO yields +11.68% from the same starting point.
• Reward distribution analysis shows high-variance samples (group mean around 0.2-0.65) barely change during GRPO training, while the low-variance region steadily grows.

Limitations

• Dr. GRPO remains susceptible to difficulty bias, as the authors acknowledge
• Gap to best absolute scores remains on NAVSIM and WaymoE2E for some reasoning-based VLAs
• Evaluated only on NAVSIM and WaymoE2E with a single 3B backbone

Experimental setup

Datasets: NAVSIM navtrain WaymoE2E

Benchmarks: NAVSIM navtest WaymoE2E test

Metrics: PDMS Collision DAC Direction Progress TTC Comfort RFS ADE@3

Baselines: Poutine HMVLM DiffusionLTF UniPlan AutoVLA UniAD Transfuser Hydra-MDP DiffusionDrive RecogDrive

Builds on: GRPO Dr. GRPO Qwen2.5-VL DAPO

Closest papers by meaning

• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.91)
• LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
• From Representational Complementarity to Dual Systems: Synergizing VLM and Vision-Only Backbones for End-to-End Driving (NeurIPS 2026 · similarity 0.89)
• All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models (CVPR 2026 · similarity 0.89)