Bring Reason to Vision: Understanding Perception and Reasoning through Model Merging

Shiqi Chen*, Jinghan Zhang*, Tongyao Zhu, Wei Liu, Siyang Gao, Miao Xiong, Manling Li, Junxian He$^\dagger$

paper |

• Abstract: We enhance VLM reasoning via model merging and understand perception and reasoning ability inside model.

Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas

Shiqi Chen, Tongyao Zhu, Ruochen Zhou, Jinghan Zhang, Siyang Gao, Juan Carlos Niebles, Mor Geva, Junxian He, Jiajun Wu, Manling Li$^\dagger$

paper | dataset |

• Abstract: A training-free decoding method called AdaptVis boosts VLM’s spatial reasoning by dynamically sharpening or broadening attention based on confidence, yielding up to 50-point accuracy gains on benchmarks.

Compression Represents Intelligence Linearly

Yuzhen Huang*, Jinghan Zhang*, Zifei Shan, Junxian He$^\dagger$

paper | leaderboard | dataset |

• Abstract: In this paper, we study the relationship between compression rate and intelligence of LLMs.

Composing Parameter-Efficient Modules with Arithmetic Operations

Jinghan Zhang, Shiqi Chen, Junteng Liu, Junxian He$^\dagger$

paper |

• Abstract: In this paper, we study model merging on parameter-efficient modules like LoRA and (IA)^3.

C-Eval: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models

Yuzhen Huang*, Yuzhuo Bai*, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, Junxian He$^\dagger$

paper | website | dataset |

FELM: Benchmarking Factuality Evaluation of Large Language Models

Shiqi Chen, Yiran Zhao, Jinghan Zhang, I-Chun Chern, Siyang Gao, Pengfei Liu, Junxian He$^\dagger$

paper | website | dataset |