MIRA-7B
This is the main Qwen2.5-VL-7B checkpoint from MIRA: Reinforcing Multimodal Reasoning via Deceptive Contextual Augmentation (NeurIPS 2026). MIRA trains a vision-language model to recover from misleading visual interpretations during reasoning.
The repository contains the complete Hugging Face-format model weights and the configuration and processor files required to load the checkpoint. Consult the official code repository for the training recipe and benchmark evaluation.
The model is a fine-tuned derivative of Qwen2.5-VL-7B-Instruct. Its Apache-2.0 license and upstream licensing information are included in this repository. Please consult the paper and source datasets for details about the training and evaluation data.