CogSense-8B

This repository contains the weights for CogSense-8B, a Multimodal Large Language Model (MLLM) introduced in the paper Toward Cognitive Supersensing in Multimodal Large Language Model.

Project Page | Code | Paper

Introduction

CogSense-8B is trained using Cognitive Supersensing, a novel training paradigm that endows MLLMs with human-like visual imagery capabilities. By integrating a Latent Visual Imagery Prediction (LVIP) head, the model learns sequences of visual cognitive latent embeddings and aligns them with answers, forming vision-based internal reasoning chains. This approach aims to bridge the gap between perceptual recognition and complex cognitive understanding.

CogSense-Bench

The model's cognitive capabilities are evaluated on CogSense-Bench, a comprehensive visual question answering (VQA) benchmark assessing five cognitive dimensions:

• Fluid intelligence
• Crystallized intelligence
• Visuospatial cognition
• Mental simulation
• Visual routines

Citation

If you find this work useful, please consider citing:

@inproceedings{li2026cognitive,
  title     = {Toward Cognitive Supersensing in Multimodal Large Language Models},
  author    = {Li, Boyi and Shen, Yifan and Liu, Yuanzhe and Xu, Yifan and
               Liu, Jiateng and Li, Xinzhuo and Li, Zhengyuan and Zhu, Jingyuan and
               Zhong, Yunhan and Lan, Fangzhou and Cao, Jianguo and Rehg, James M. and
               Ji, Heng and Lourentzou, Ismini and Cao, Xu},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}