Stanford paper: more info helps some LLM trading agents, wrecks others
rohanpaul_ai · x · 2026-11-08
A Stanford team's arXiv paper, "Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies," tests 10 frontier LLMs across multi-resource trade, bilateral bargaining, and double auctions under varying disclosure precision.
Key finding: more information is not always better, and effects are highly model-dependent:
- In trade, revealing exact valuations cut GPT-6 Sol's gains from 98% to 91% of maximum, but raised Claude Sonnet 5's from 72% to 86%
- In auctions, full disclosure inflated price demands, dropping Grok 4.7 from 97% to 62% and Claude Opus 5.5 from 94% to 64%
- In bargaining, exact info made Grok 4.7's agreements less balanced, while Opus 5.5 kept splitting gains symmetrically
Conclusion: model selection and information design must be co-optimized — when agents negotiate for you, test the model and what it sees together.
Related event: Stanford Paper: More Information Can Hurt LLM Trading Agents(2 posts)→
More from coding & agent
- Dutch-Language Bug: A 15k-Char Prompt Made Devs Blame the Model for Their Own Mistake — Initial_Orange2985 · 2026-11-08
- zml team hits "RSI": DeepSeek 4.1 Flash token use exceeds Codex+Claude combined — RemiCadene · 2026-11-08
- Agent S4 lets agents evolve reusable computer-use interfaces from their own task experience — burkov · 2026-11-08
- Running 50-100 Agents, Dev Weighs Local Decision Models to Cut Latency — f1yingbanana · 2026-11-08
- This PM replaced the PRD with an interactive Claude artifact — PawelHuryn · 2026-11-08
- Spacedrop lets you drag-and-drop code to deploy in seconds, no GitHub push needed — saheedniyi_02 · 2026-11-08