hzxllll/DEFT-RLVR-model-HF

VISIONPricing:Input $0.727 / Output $5.405Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The hzxllll/DEFT-RLVR-model-HF is an 8 billion parameter vision-language model adapted from Qwen3-VL-8B-Instruct by hzxllll. It utilizes Deferred Exposure of Future Trajectories for Verifiable Reasoning (DEFT-RLVR) to address trajectory anchoring bias in autonomous driving. This model is specifically designed for research in autonomous-driving visual reasoning, candidate-grounded decision making, and reinforcement learning with verifiable rewards, processing multi-view driving visual context and text instructions.

Loading preview...

Overview of DEFT-RLVR Model

The hzxllll/DEFT-RLVR-model-HF is an 8 billion parameter vision-language model (VLM) built upon the Qwen3-VL-8B-Instruct architecture. Developed by hzxllll, this model incorporates the Deferred Exposure of Future Trajectories for Verifiable Reasoning (DEFT-RLVR) method, specifically tailored for autonomous driving applications.

Key Capabilities and Innovations

  • Addresses Trajectory Anchoring Bias: The model is designed to mitigate the issue where VLMs rationalize known outcomes rather than inferring causally faithful decisions from scene evidence in autonomous driving reasoning tasks.
  • DEFT-RLVR Methodology: This approach requires the VLM to commit to a high-level driving decision based on scene evidence before candidate trajectories are revealed, promoting more robust and verifiable reasoning.
  • AD-MCQ Dataset Integration: The model was adapted using the AD-MCQ dataset, which formulates planning as a verifiable selection among explicit candidate trajectories.
  • Input/Output: It processes multi-view driving visual context and text instructions, generating scene-grounded reasoning and candidate selections.

Intended Use Cases

This checkpoint is primarily intended for research purposes in several key areas:

  • Autonomous-driving scene understanding and causal reasoning.
  • Candidate-trajectory selection in complex driving scenarios.
  • Multimodal reasoning evaluation within the autonomous driving domain.
  • Reinforcement Learning with Verifiable Rewards (RLVR) and process-supervised policy adaptation.

Finetunes
1 models