The paper introduces StreamDuo, a live game commentator that calls on a research agent to look up wiki entries when it hits something it can't name.
Setup: A trainable commentator pairs with a frozen agent that reads the game's wiki. The entries come back about two seconds later, while the stream carries on.
Training: Only the commentator is trained, and only on Baldur's Gate 3.
Results: Across ten unseen games, entity recall rises from 0.082 to 0.116 over Proact-VL trained on the same data, and source-game leakage drops from 9.9% to 1.0%.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper introduces StreamDuo, a live game commentator that calls on a research agent to look up wiki entries when it hits something it can't name.