hzxllll/DEFT-RLVR-model-HF

DEFT-RLVR Model

This repository contains the Qwen3-VL-8B checkpoint from Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs.

The model is adapted from Qwen3-VL-8B-Instruct using Deferred Exposure of Future Trajectories for RLVR (DEFT-RLVR) on our AD-MCQ dataset.

Overview

Autonomous-driving reasoning supervision often reveals the logged ground-truth future trajectory before asking a vision-language model to explain its decision. Our paper identifies this as trajectory anchoring bias: the model can rationalize a known outcome instead of inferring a causally faithful decision from scene evidence.

We introduce two components:

• AD-MCQ formulates planning as an exactly verifiable selection among explicit candidate trajectories.
• DEFT-RLVR requires the policy to reason about the scene and commit to a high-level driving decision before candidate trajectories are revealed for grounding and verification.

This checkpoint is intended for research on autonomous-driving visual reasoning, candidate-grounded decision making, and reinforcement learning with verifiable rewards.

Resources

• Paper: Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs
• Code: github.com/hzx122/DEFT-RLVR
• Dataset: hzxllll/AD-MCQ
• Base model: Qwen/Qwen3-VL-8B-Instruct

Model Details

Item	Description
Architecture	Qwen3VLForConditionalGeneration
Base model	Qwen3-VL-8B-Instruct
Training method	DEFT-RLVR
Training task	Candidate-trajectory multiple-choice reasoning
Input	Multi-view driving visual context and text instructions
Output	Scene-grounded reasoning and candidate selection
License	Apache 2.0

Usage

Install a Transformers version that supports Qwen3-VL, then load the checkpoint with the standard Hugging Face API:

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "hzxllll/DEFT-RLVR-model-HF"

processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/driving_frame.jpg"},
            {"type": "text", "text": "Describe the driving scene and determine the appropriate high-level driving decision."},
        ],
    }
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=512)

generated_ids = generated_ids[:, inputs.input_ids.shape[1]:]
print(processor.batch_decode(generated_ids, skip_special_tokens=True)[0])

For the paper's two-turn DEFT protocol and AD-MCQ candidate format, use the prompts and evaluation code provided in the project repository.

Intended Use

The checkpoint is released for research in:

• autonomous-driving scene understanding and causal reasoning;
• candidate-trajectory selection;
• multimodal reasoning evaluation;
• RLVR and process-supervised policy adaptation.

Citation

@article{huang2026deftrlvr,
  title        = {Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs},
  author       = {Huang, Zixuan and Zhou, Yang and Wang, Kaixuan and Zhang, Guli and Xie, Hongyan and Zhu, Yakun and Geng, Hao and Ban, Yikun and Wang, Deqing},
  year         = {2026},
  journal      = {arXiv preprint arXiv:2608.01755},
  url          = {https://arxiv.org/abs/2608.01755}
}

Acknowledgements

This model is built on Qwen3-VL-8B-Instruct. We thank the Qwen team and the open-source community for their contributions.