Zefan Cai
@Zefan_Cai
Followers
211
Following
159
Media
18
Statuses
117
Now Ph.D student @UWMadison Previous @PKU1898
Joined May 2023
RT @Xinyu2ML: 🚀 Super excited to share Multiverse!. 🏃 It’s been a long journey exploring the space between model design and hardware effici….
0
20
0
I feel that the long context is the final important question. Once you could put all the training data used in SFT into context and still keep the inference cost constant, we could solve most of the current problems.
0
0
2
If you feel our work useful, please site us:. @misc{huang2025vista,. title={VisualToolAgent (VisTA): A Reinforcement Learning Framework for Visual Tool Selection}, . author={Zeyi Huang and Yuyang Ji and Anirudh Sundara Rajan and Zefan Cai and Wen Xiao and Junjie Hu and.
0
0
0
If VisTA sparks ideas for adaptive tool-using agents, drop a ⭐ on the repo and cite us: 2505.20289. Let’s build vision systems that choose wisely 🧠⚡️ #AI #Vision #ReinforcementLearning #LLM #OpenSource.
0
0
1
Tech bits 🛠️.• Discrete action space = tool IDs.• Reward = downstream task score.• GRPO stabilizes updates across tool groups.• Training: 4×A100, 1-day synth RL rollouts.
0
0
0
Plug & Play ⚙️.Just wrap your favorite vision-LLM: GPT-4o, Qwen-VL, whatever. VisTA drives the tool calls, the LLM focuses on reasoning.
0
0
0
Generalization FTW 🚀.On unseen tasks, VisTA still outperforms prompt-engineering & finetune baselines. Learning to explore > memorizing fixed recipes.
0
0
0
🤖 During inference, VisTA dynamically picks specialized modules: OCR for charts, geometric solvers for diagrams, etc. The policy remains model-agnostic—swap the underlying LLM and keep the same tool selector.
0
0
0
Given a user query, the agent selects tools from a pre-defined set of external tools. The tools are applied to the image, and their outputs and the query are fed.to a frozen reasoner model.
0
0
0
Core idea: treat tool choice as an action. VisTA explores a library of visual modules, gets task feedback, and learns a policy with Group Relative Policy Optimization (GRPO)—zero extra supervision needed.
0
0
0
Meet VisTA (VisualToolAgent): a RL framework that teaches vision models to PICK the right tool for every visual-reasoning task—no handcrafted prompts, no human demos. Paper 👉
arxiv.org
We introduce VisTA, a new reinforcement learning framework that empowers visual agents to dynamically explore, select, and combine tools from a diverse library based on empirical performance....
0
0
1
Project Page: ARXIV: Code:
github.com
VisualToolAgent (VisTA): A Reinforcement Learning Framework for Visual Tool Selection - OoDBag/VisTA
0
0
1
RT @huang43602: 🚨Our new paper: VisualToolAgent (VisTA) 🚨. Visual agents learn to use tools—no prompts or supervision! . ✅RL via GRPO.✅Dec….
0
5
0
Really appreciate the support for VisualToolAgent (VisTA): our new RL-based framework for dynamic tool selection in visual reasoning.
VisualToolAgent (VisTA). A Reinforcement Learning Framework for Visual Tool Selection
18
9
59
Super excited to work with my amazing collaborators @wendyxiao06091 @preminstrel @ChengLuo_lc @liyucheng_2 @zhendongucb @AnimaAnandkumar @JunjieHu12.
0
0
1
If R-KV helps your research, please cite 🎓.BibTeX & full details in the repo. Let’s make long-form reasoning efficient together! #AI #LLM #KVCache #MLOps #OpenSource #MLSys #SparseAttention.
0
0
0
Everything’s open-source: code, scripts, Jupyter analysis notebooks. 🚀 Fork it, test it, file issues—we’d love feedback. GitHub:
github.com
R-KV: Redundancy-aware KV Cache Compression for Reasoning Models - Zefan-Cai/R-KV
0
0
0