Jize Jiang

CS PhD @ UIUC

[Mingyuan Wu @MingyuanWu4]

New preprint from Meta and UIUC for LLM spreadsheet agents.

We build and open-source the full stack: data, environment, verifier/reward, and RL pipeline.

🍎Project Page: https://t.co/RwgAO0SfBW
Paper: https://t.co/6rQywsHyYB
Code: https://t.co/XcnJUQwL5d

We present Spreadsheet-RL, a reinforcement learning (RL) fine-tuning framework designed to train specialized spreadsheet agents within a realistic Microsoft Excel environment.
#UIUC, #Meta, #LLM, #RL, #agent, #spreadsheet, #AI

[Jize Jiang @JizeJiang]

Big thanks to everyone on the team and our mentors 🌟

I’m thrilled that PlugMem has been accepted to ICML 2026. This is a big milestone for our work on memory for evolving agents.

What excites me just as much is we are turning PlugMem into something people can actually build with, a truly plug-and-play memory module that works across real agent runtimes and interpretable through visualization interfaces.

Making research accessible is part of pushing the frontier 🎆

[Ke Yang @EmpathYang]

Big PlugMem update 🧠
A plug-and-play memory module for LLM agents — turns raw trajectories into a knowledge graph your agent actually reasons over.
🎉 Accepted to ICML 2026
🔌 Drop it into OpenClaw 🦞, Claude Code, and other agent runtimes
🔍 Visualize memory · test retrieval · replay sessions
🥇 SOTA backbone on LongMemEval & HotpotQA — general enough to build on
Paper: https://t.co/9IKkUUYeaV
Code: https://t.co/3HlfVzWPgZ
#ICML2026 #LLM #Agents

[Microsoft Research @MSFTResearch]

PlugMem transforms AI agents’ interaction histories into structured, reusable knowledge. It integrates with any agent, supports diverse tasks and memory types, and maximizes decision quality while significantly reducing memory token use: https://t.co/girJeCrr6p

[Ke Yang @EmpathYang]

📰New preprint: How can we build a task-agnostic plug-and-play memory module for LLM agents that supports multiple memory types?
We present PlugMem🔌🧠, a plugin memory module that works across tasks by turning heterogeneous experience into knowledge.
Evaluated unchanged on long-term dialogue🗣️, multi-hop QA🕵️, and web agents🕸️🤖, PlugMem improves performance while using far fewer memory tokens.
📜Paper: https://t.co/A8tNQjkCCb
🔨Code: https://t.co/mt1aJKxQIz

[Ke Yang @EmpathYang]

We’ve been building a task-agnostic memory module for LLM agents — PlugMem.

While running experiments across long-horizon QA, multi-hop retrieval, and web agents, we found several unexpected patterns about how memory actually helps (or hurts) decision-making.

Code: https://t.co/mt1aJKxQIz

Work with an amazing team: @chenzixi23, @XuanHe21, @JizeJiang, the deep learning group @MSFTResearch, @dmguiuc, and @TIMANUIUC.
Thread ↓

[Mingyuan Wu @MingyuanWu4]

Can VLMs learn to reason better by drawing on the brilliant thoughts of others.

🔥Our recent work on vision language model reasoning, through carefully designed multimodal memory and retrieval, has been accepted to Main Conference of #EMNLP2025.

💡Inspired by case-based reasoning, we introduce Cache of Thought, a dynamic memory that stores high-quality answers and thoughts from master VLMs. This cache serves as guidance for apprentice VLMs, helping them generate stronger responses to similar multimodal queries.

🍕Enjoy the best of accuracy and efficiency with joint inference framework that integrates VLMs of different sizes.

📝Paper: https://t.co/cBUgBGzINa

This work was created in long-term collaboration with amazing friends. We brainstormed the idea together, split the implementation, sought advice and feedback from our advisors, and spent late nights writing before the deadline. We would like to call this style “research with friends.”

We hope to keep it alive until the very end of our PhD journeys, so that we can also keep learning to reason better by drawing on the brilliant thoughts of others.

[Jize Jiang @JizeJiang]

Excited to introduce VTool-R1!

We’ve trained VLMs to “think visually” using RL, blending Python-based 🖼️visual edits with💡textual Chain-of-Thought reasoning.
Our trained qwen2.5-VL-32B surpasses GPT-4o on ChartQA & TableVQA, and even the compact qwen2.5-VL-7B significantly narrows the gap.

💭web: https://t.co/1j3gdqYDxn
📝paper: https://t.co/3oIWTj4wVv
🌟code: https://t.co/VzIs1TlYU6

This is a work done collaboratively between some amazing friends.
Special thanks to @BowenJin13 for advise and @XingyuFu2 for data and tool sets. Also thank all the open sourced infras. 🙏