Stanford Paper Shows Full Disclosure Can Hurt LLM Trading Agents Across 10 Frontier Models
shangbinfeng · x · 2026-11-13
A Stanford/University of Maryland/All Purpose AI study, "Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies," tests how information policies shape 10 frontier LLM agents in strategic markets.
Across three settings, comparing coarse vs. full disclosure (change in % of maximum gains):
- Multi-resource trade: full disclosure helps Grok 4.3 (+26.9), Sonnet 5 (+13.5), GPT-6 Luna (+10.1), but hurts GPT-5 mini (-10.2) and GPT-5.1 (-8.8).
- Single-good bargaining (normalized Nash score): most models gain with full disclosure; GPT-5 mini (-0.46) and GPT-5.1 (-0.20) lose.
- Double auctions: full disclosure is disastrous for most — Grok 4.7 (-35.3), Opus 5.5 (-30.6), Gemini 3.1 Pro (-18.4); only Gemini 3.7 Flash benefits (+1.9).
Key takeaway: disclosure can help agents find better trades or push them to demand a larger share and break negotiations — how much information to give agents is itself a design decision that varies by model and market structure.
More from AGI Musings
- Analyst: Everyone Underestimates Non-AI R&D Acceleration — It'll All Be Ripping Anyway — teortaxesTex · 2026-11-13
- Sentience Institute validates a psychometric scale for 'substratism' — moral bias against AI — jacyanthis · 2026-11-13
- Anthropic's alignment work paying off as OpenAI falls behind, argues researcher — haider1 · 2026-11-13
- Claim: Senior researchers at OpenAI and Anthropic have confirmed contested study result — NathanpmYoung · 2026-11-13
- Investor tszzl: Chinese open-source models have earned huge global goodwill and soft power — tszzl · 2026-11-13
- 'We Have a Right to Compute': Critics Warn US AI Regulation Follows Nuclear Energy's Playbook — nptacek · 2026-11-13