The Great Library

too lazy to edit :)

TODO: 一句话概括这篇论文

This work focuses on 3 things: 1. in Agent's react loop, ask VLM rollout K candidate actions for next step, then ask Judge select most proximate one. 2. This work has trained model by GRPO with single `click reward` which outperforms other models. 3. Data cleanup: this work uses many open-source grounding training data to train the grounding model. The origin data is scraped by programs which is cheap but may contain errors. The team uses OmniParser to cleanup the origin training data.

This work introduces a grounding benchmark OSWorld-G and a training dataset JEDI. Tasks of OSWorld-G come from trajectories of OSWorld. Training data of JEDI is synthesized.

Agent S2 uses an expert model for grounding and replans once each subgoal is completed.

Paper Summary of "VISTA: A Visual Harness for Reasoning in an Interactive World".

Paper Summary of "Agent S: An Open Agentic Framework that Uses Computers Like a Human".

The hardware and software configuration of my laptop.