Update README.md

- AICO (Action-centric Interaction Chain Optimization), which uses only selective token masking, outperforms TRICO on simple tasks

- TRICO demonstrates superior exploration capabilities on more complex problems

<img width="990" alt="image" src="./public/1.png" />

<img width="800" alt="image" src="./public/1.png" />

<img width="990" alt="image" src="./public/2.png" />

<img width="800" alt="image" src="./public/2.png" />

<img width="990" alt="image" src="./public/3.png" />

<img width="800" alt="image" src="./public/3.png" />

## Comparison of Algorithms

![image (7)](https://github.com/user-attachments/assets/ddea7352-0a14-45a5-94a9-655a07c9fe3e)

# Project Roadmap

 ## Limitations and Future Work

## Short-term Goals

- [ ] Address training stability issues

  - [ ] Implement improved early stopping techniques

  - [ ] Investigate root causes of instability

  - [ ] Test alternative optimization strategies

- [ ] Expand evaluation framework to more diverse visual environments

- [ ] Scaling to larger models and applying TRICO to text-only tasks

- [ ] Merge to RAGEN for better package mangement

- Training can be unstable, often requiring early stopping

- We aim to expand evaluation to more diverse visual environments

- Future plans include scaling to larger models and applying TRICO to text-only tasks

## Acknowledgement

We thank [RAGEN](https://github.com/RAGEN-AI/RAGEN) for its innovative exploration in multi-turn reinforcement learning for LLM agents. We thank [verl](https://github.com/volcengine/verl) for its RL framework. We thank [EasyR1](https://github.com/hiyouga/EasyR1) for adding initial support for VLMs to verl.