EA-RAM: Auction-based LLM routing cuts costs, beats centralized routers

Routing queries to the right large language model (LLM) is a cost-quality balancing act, but most routers rely on a centralized predictor that estimates model performance before sending a query. That creates an information-risk mismatch—the router doesn't know each provider's actual capabilities—and doesn't scale as the pool of models grows. A new paper (arXiv:2608.12719) from Haolong Chen and colleagues flips the paradigm: instead of predicting quality centrally, they run a reverse auction where LLM providers bid on each query with their own self-assessed success probabilities and execution costs.

The proposed mechanism, called Error-Aware Reverse Auction Mechanism (EA-RAM), explicitly models the inherent "Dual Error"—noise in both provider self-predictions and center evaluations. The authors prove EA-RAM is Bayesian incentive compatible and individually rational even under this noise, meaning providers are incentivized to bid honestly and won't lose money by participating. They also derive a welfare-loss bound and identify robustness effects: opposite-signed errors can cancel each other out, saturating link functions like logistic smooth extreme cases, and extra noise flattens belief maps, reducing the benefit of gaming the system. Experiments on simulations and real-world benchmarks show EA-RAM achieves a better cost-performance Pareto frontier than centralized baselines, and performance improves further when providers contribute local information. The work suggests a future where LLM routing isn't a centralized brain but a transparent, market-driven marketplace—and it could change how enterprises select models at scale.