For instance, within the state of affairs illustrated within the determine above, the shared sequence of tokens could be “Mechanism Design for”. The distributions could be [(“Large”, 0.8), (“Generative”, 0.2)] for the LLM of Agent 1, (“Giant”, 1.0) for the LLM of Agent 2, and (“Generative”, 1.0) for the LLM of Agent 3. The bids could be 1, 2, and a couple of, respectively. A attainable aggregated distribution could be the bid-weighted common of the distributions, particularly [(“Large”, 0.56), (“Generative”, 0.44)]. A attainable alternative for the funds could be to ask every agent to pay their bid, which might have the brokers commit 1, 2, and a couple of, respectively.
For our theoretical evaluation of this mannequin (and alternatives of distribution aggregation features and fee features), we assume that the brokers in truth report their distributions, however could also be strategic about their bids. We consider it is a life like assumption, as LLMs encode preferences over output textual content in a succinct and non-obvious approach. Furthermore, to ensure that the token public sale to have the ability to mixture distributions, we have to have (at the least) some (minimal) details about agent’s preferences away from their “most well-liked” distributions. Our method right here is to imagine that the brokers have (identified) partial desire orders over distributions. That’s, we assume that brokers could possibly rank some, however not all, pairs of distributions.