LLM agents outperform humans in matching market stability and truthfulness
New arXiv study finds LLM agents report preferences truthfully at higher rates than humans in matching markets
A new paper on arXiv (2606.03030) by Yukihiro Hoshino, Ayato Kitadai, and Nariaki Nishino investigates whether classical matching mechanisms—like Deferred Acceptance (DA), Efficiency-Adjusted DA (EADA), and Top Trading Cycles (TTC)—function as intended when LLM agents act as delegated decision-makers in allocation markets. The researchers compared decentralized free-negotiation markets with centralized mechanism-based markets across controlled one-to-one matching environments. Their results show that mechanism-based markets consistently outperform free negotiation in both stability and efficiency.
Interestingly, LLM agents reported their preferences truthfully at substantially higher rates than human subjects in comparable DA and EADA settings. However, truth-telling was not uniformly aligned with formal strategy-proofness: TTC, despite being theoretically strategy-proof, did not always elicit higher truth-telling than EADA. The authors conclude that matching theory provides a useful but incomplete guide for designing institutions in LLM-agent markets, opening new questions about how to design markets for AI agents.
- Mechanism-based markets (DA, EADA, TTC) outperform free-negotiation markets in stability and efficiency when LLM agents make decisions
- LLM agents report preferences truthfully at substantially higher rates than human subjects in DA and EADA environments
- TTC, though strategy-proof, does not always elicit higher truth-telling than EADA, showing strategy-proofness alone is insufficient
Why It Matters
As AI agents handle more economic decisions, understanding how matching mechanisms govern them is critical for market design