📷 Selected Publication

MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory

TL;DR: MemEye is a vision-centric long-term memory benchmark that evaluates agents’ ability to remember, update, and reason over visual information across long-running, multi-session image-grounded interactions.

Zeru Shi*, Minghao Guo*, Qingyue Jiao*, Yihao Quan, Boxuan Zhang, Danrui Li, Liwei Che, Wujiang Xu, Shilong Liu, Zirui Liu, Mubbasir Kapadia, Vladimir Pavlovic, Jiang Liu, Mengdi Wang, Yiyu Shi, Dimitris N. Metaxas, Ruixiang Tang

[Project Page]   [Arxiv]  [GitHub]

From Commands to Prompts: LLM-based Semantic File System

[ICLR 2025]

TL;DR: We propose a vector-based agent memory system that enables users to manage and interact with computer files through natural language, eliminating the need for traditional Linux commands.

Zeru Shi*, Kai Mei*, Mingyu Jin, Yongye Su, Chaoji Zuo, Wenyue Hua, Wujiang Xu, Yujie Ren, Zirui Liu, Mengnan Du, Dong Deng, Yongfeng Zhang

[Arxiv] [GitHub]

Castle: Causal Cascade Updates in Relational Databases with Large Language Models

[EMNLP 2025, Main]

Yongye Su, Yucheng Zhang, Zeru Shi, Bruno Ribeiro, Elisa Bertino

[Arxiv]

Online Auditing for Early Failure Prediction in Multi-Agent Systems

Boxuan Zhang *, Jianing Zhu *, Zeru Shi, Dongfang Liu, Ruixiang Tang

[Arxiv]

A Single Layer to Explain Them All: Understanding Massive Values in Large Language Models

[ICML 2026]

Zeru Shi, Zhenting Wang, Fan Yang, Qifan Wang, Ruixiang Tang

TL;DR: We explored the actionable mechanistic interpretation of massive values in LLMs and propose a method to mitigate the massive activatons.

[Project Page]   [Arxiv]  [GitHub]

Meaningless Tokens, Meaningful Gains: How Activation Shifts Enhance LLM Reasoning

Zeru Shi, Yingjia Wan, Zhenting Wang, Qifan Wang, Fan Yang, Elisa Kreiss, Ruixiang Tang

[Arxiv], [GitHub]

Improving Visual Reasoning with Iterative Evidence Refinement

Zeru Shi*, Kai Mei*, Yihao Quan, Dimitris N. Metaxas, Ruixiang Tang

[Arxiv],

SeFENet: Robust Deep Homography Estimation via Semantic-Driven Feature Enhancement

[IEEE TCSVT, IF=11.1]

TL;DR: We design a meta-learning framework to improve the robustness and performance of homography estimation under challenging environments.

Zeru Shi, Zengxi Zhang, Kemeng Cui, Ruizhe An, Jinyuan Liu, Zhiying Jiang

[Arxiv]