Stanford Paper Finds More Information Can Hurt Some LLM Trading Agents

A new Stanford paper tested 10 frontier LLM agents in trading, bargaining and double auctions, varying only the precision of disclosed private valuations. More information was not universally beneficial: exact values cut GPT-6 Sol's gains from 98% to 91% of maximum but raised Claude Sonnet 5's from 72% to 86%, while full disclosure in auctions inflated price demands and dropped Grok 4.7 from 97% to 62%. The takeaway is that multi-agent systems need model-specific rules on what each agent can see.