More shared data cut market gains for some AI models, not all

Researchers tested if more information helps AI agents when they trade. They used 10 frontier language models in trade, bargaining and double auction markets. The results are different for each model. More information increased gains for Claude Sonnet 5 in trade, from 72% to 86%. It decreased them for GPT-6 Sol, from 98% to 91%.

How it works

The authors name the method empirical information design. They keep the rules and the models the same and change only the data that each agent sees. They compare results with the actual values of each deal.

The numbers

In double auctions, full disclosure made price demands higher and caused a market stop. Gemini 3.1 Pro decreased from 97% to 78%. Grok 4.7 decreased from 97% to 62%. In bargaining, correct data made Grok 4.7 produce less balanced deals. Claude Opus 5.5 divided the gains equally.

What is not known

The paper does not give a rule that tells which model will get more. The authors tell only that the selection of a model and the design of the information must be tuned together.

Insight: A change of model can break the data rules around it

Cohere made Rerank 5 models that give an LLM only the parts of a file that are about the query. This paper shows that more information can decrease results, and that the effect is different for each model. The data that each agent sees is at this time tied to the model, and a change of model can break a setup that did work. Test the data rules again after each change of model.