AIBRIEF
A replayable market simulator lets LLM agents test interventions in a "lab" before a crisis decision. On one deployment (Luna) — the only one of three with a complete record — lab access cut regret 57%, but the caveats are large.
A new paper introduces CrashGym, a deterministic market simulator asking a narrow question: does letting an LLM agent run practice interventions in a "lab" copy of a market actually make its crisis decision better?
The agent faces a market about to suffer a shock-triggered crash and chooses between two interventions: deleveraging, which forces leveraged traders to close positions, or a liquidity backstop, standing ready to buy. Before committing, it can run both in a practice realization, then is scored on a separately seeded target. The score is regret, the gap between the chosen outcome and the best intervention computed by replaying both options. Four switchable amplification channels (persistent price impact, liquidity withdrawal, margin-driven liquidation, an evolving strategy mix) let researchers stage how realistic the practice world is.
The headline result comes from Luna, the only one of three prospectively specified deployments with a complete record. Lab access cut mean regret from 8.878 to 3.785 utility units, a 57.4% reduction (paired reduction 5.092, nominal 95% world-bootstrap interval 1.028 to 9.507). Two other prospective deployments are incomplete; an exploratory study of four deployments shows the same direction across all four.
The paper also reports a negative result: a lab whose strategy mix evolves endogenously offers no established advantage over a fixed-mix lab. The lesson is that how the practice world is staged matters, not just having one.
CrashGym is an arXiv preprint, not peer-reviewed.