Language Models Can Reduce Asymmetry in Information Markets (COLM 2024)

I was a co-lead author with Nasim Rahaman (plus Yoshua Bengio, Manuel Wüthrich, Li Erran Li, Bernhard Schölkopf and Chris Pal). Sharing because this work explains more of what I see in the market.

Arrow's Information Paradox (1962) explains why information markets are inefficient. It says: you need to see information in order to choose whether to buy it, but once you've seen it you no longer need to buy it [https://en.wikipedia.org/wiki/Arrow_information_paradox].

LLMs move the pareto frontier. To show this, we set up a system where buyer agents post queries and vendors reply with priced quotes. Buyer agents can actually read the quote, and the price, and their own value function to make a decision in the context of all the information. And if they choose not to buy, then the information is simply erased from their memory.

This was also the paper where we figured out that debate prompting squeezed performance out of LLMs when making rational and economic decisions. This was in the pre-reasoning model era, and before the ICML best paper that investigated this phenomeon more deeply.

Code (research code, MIT): https://github.com/tn-learn/info-bazaar Arxiv: https://arxiv.org/pdf/2403.14443 Lacuna: https://lacuna.tiptreesystems.com/work/redesigning-informati...