StreamDuo for cross-game streaming commentary
Original title: StreamDuo: A Live Commentator and a Research Agent for Unseen Games
StreamDuo pairs a commentator with a research agent using game wikis to enable commentary on unseen games.
The paper studies cross-game streaming commentary where models trained on one game commentate on others. StreamDuo uses a trainable commentator and a frozen research agent that reads the wiki. Trained on Baldur's Gate 3, it improves entity recall and reduces source-game leakage on ten unseen games.
- Raises entity recall from 0.082 to 0.116
- Cuts source-game leakage from 9.9% to 1.0%
- Trained on one game in two stages
- Tested on ten unseen games
- Achieves best timing and judged quality
- Tags
- streaming
- Organizations
- Hugging Face
- AI models
- StreamDuo
From the original article
1 min read at huggingface.co
Streaming video LLMs can now commentate live gameplay, but they are trained and tested on the same games. We study cross-game streaming commentary, in which a model trained on one game commentates games it has never seen, and find that current commentators transfer their timing but not their knowledge of the game world.
By Zixian Gao, Penghao Luo, Xiangcheng Cao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.