2 ms·

I plan to read the paper later, but anyone have a TL;DR for how they connected a language model to the game state? That seems like the real advance here. Langua

I plan to read the paper later, but anyone have a TL;DR for how they connected a language model to the game state? That seems like the real advance here. Language models are so prone to making stuff up and spouting nonsense, and controlling them is really hard.

LucasBrandt 4y ago
> At a high level, Cicero combines a dialogue module with a strategic reasoning module, along with a filtering process to reject low-quality messages. The strategic reasoning module generates "intents" (a set of planned actions for the agent and its speaking partner) which are sent to the dialogue module. They trained the model to be able to generate messages from those intents using data from messages in real games which were manually labeled with intents based on the content of the message.
- sanxiyn 4y ago
  No, intent labeling wasn't manual, it was done automatically from raw data.