OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

📄 Paper (arXiv) | 💻 GitHub | 🌐 Project Page

Xiaomi Embodied Intelligence Team

---

Overview

OneVL is a Vision-Language-Action (VLA) framework for autonomous driving that achieves state-of-the-art trajectory prediction accuracy while matching the inference latency of answer-only autoregressive models.

Prior latent Chain-of-Thought (CoT) methods compress reasoning into opaque hidden states — fast, but consistently underperform explicit CoT on driving tasks. OneVL identifies the root cause: purely linguistic latents encode abstract semantic labels rather than the spatiotemporal causal dynamics that govern real driving scenes. OneVL addresses this with dual-modal auxiliary decoders that force compact latent tokens to encode both human-readable reasoning and future scene dynamics simultaneously.

At inference, both decoders are discarded and all latents are prefilled into the prompt context in a single parallel pass — matching answer-only AR prediction speed while recovering the interpretability of explicit CoT in both vision and language.

OneVL is the first latent CoT method to surpass explicit autoregressive CoT across all four driving benchmarks.

---

Architecture

OneVL augments Qwen3-VL-4B-Instruct with three components:

Latent Token Interface — 4 visual latent tokens + 2 language latent tokens are inserted in the assistant response before the answer, using existing vocabulary tokens (no new special tokens added).

Visual Auxiliary Decoder — Predicts future-frame visual tokens at t+0.5s and t+1.0s from visual latent hidden states (using the Emu3.5 IBQ 131k codebook). Acts as a world model supervision signal that forces the latent space to encode genuine physical scene dynamics — agent trajectories, road geometry, and environmental change — rather than abstract descriptions.

Language Auxiliary Decoder — Reconstructs explicit CoT reasoning text from language latent hidden states, conditioned on ViT visual features. Recovers 97% of explicit CoT text quality while running at answer-only speed.

Prefill Inference — Both decoders are discarded at inference time. All latent tokens are processed in a single parallel prefill pass; only the trajectory answer is generated autoregressively. This achieves 1.5× speedup over explicit CoT on NAVSIM and 2.3× on ROADWork.

Three-Stage Training Pipeline

Training proceeds in three stages to ensure stable joint optimization:

• Stage 0: Main model warmup (trajectory prediction)
• Stage 1: Auxiliary decoder warmup (language + visual decoders independently)
• Stage 2: Joint end-to-end fine-tuning (all components together)

Staged training is essential — ablation shows that skipping it collapses PDM-score from 88.84 to 67.13.

---

Results

NAVSIM

ROADWork

Impromptu

APR1 (Alpamayo-R1)

CoT Text Quality (NAVSIM)

OneVL's language auxiliary decoder recovers 97% of explicit CoT quality at answer-only inference speed.

---

Usage

Requirements

• Python 3.10+, CUDA GPU (≥16 GB VRAM recommended)
• transformers >= 4.57.0 (required for Qwen3VLForConditionalGeneration)

uv venv venv/onevl --python 3.12
source venv/onevl/bin/activate
pip install -r requirements.txt

Inference (Trajectory Prediction Only)

python infer_onevl.py \
    --model_path /path/to/OneVL-checkpoint \
    --test_set_path test_data/navsim_test.json \
    --image_base_path "" \
    --output_path output/navsim/results.json \
    --device cuda:0 \
    --num_latent 2 --num_latent_vis 4 \
    --max_new_tokens 1024 --answer_prefix "[" --prefix_k 0

Inference with Language + Visual Explanation

python infer_onevl.py \
    --model_path /path/to/OneVL-checkpoint \
    --test_set_path test_data/navsim_test.json \
    --image_base_path "" \
    --output_path output/navsim/results_explain.json \
    --device cuda:0 \
    --num_latent 2 --num_latent_vis 4 \
    --max_new_tokens 1024 --answer_prefix "[" --prefix_k 0 \
    --decoder_explain --aux_visual_condition \
    --c_thought 2 --max_explain_tokens 1024 \
    --visual_decoder_explain --visual_aux_visual_condition \
    --c_thought_visual 4 --max_visual_tokens 2560

Multi-GPU Inference

export MODEL_PATH=/path/to/OneVL-checkpoint
export TEST_SET_PATH=test_data/navsim_test.json
export OUTPUT_PATH=output/navsim/navsim_results.json
bash run_infer.sh

Per-benchmark scripts are available in scripts/:

bash scripts/infer_navsim.sh
bash scripts/infer_ar1.sh
bash scripts/infer_roadwork.sh
bash scripts/infer_impromptu.sh

For full documentation, evaluation scripts, and data format details, see the GitHub repository.

---

Open-Source Status

---

Citation

@article{lu2026onevl,
  title={OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation},
  author={Lu, Jinghui and Guan, Jiayi and Huang, Zhijian and Li, Jinlong and Li, Guang and Kong, Lingdong and Li, Yingyan and Wang, Han and Xu, Shaoqing and Luo, Yuechen and others},
  journal={arXiv preprint arXiv:2604.18486},
  year={2026},
  url={https://arxiv.org/abs/2604.18486}
}

---

License

Released under the Apache 2.0 License.

Model weights are built on Qwen3-VL-4B-Instruct and the visual tokenizer is from Emu3.5-VisionTokenizer; please refer to their respective licenses as well.