The Great Library

too lazy to edit :)

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 OminiParser 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.

Recoding my study & work retrospective.

GPU server: environment info, login SOP and reverse-tunnel notes.

Describe the relationship between OS, Shell and Terminal, using an example of 'What actually happens between a keystroke and a disk read.'

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

This article concludes the workflow of building a Hexo-based blog.