When Disclosure Is Compulsory: Evaluating Ride-Hailing Rebalancing Policies with Self-Interested LLM Driver Agents

It compares two families of ride-hailing rebalancing policy once their rules are public: a reactive policy (R) that sends idle drivers to cover requests just lost, and an anticipatory policy (A) that plans moves on a demand forecast. Each driver is an LLM agent maximizing its own net income, tested with no rules (N) or with rules at the directive's level of detail (L).

Two LLMs are tested as self-interested driver agents: Claude Haiku 5.5 and Jev. The paper reports that both understand the disclosed rules equally well, so the comparison of policies is not driven by a difference in rule comprehension.

The directive, to be transposed by December 2026, requires ride-hailing platforms to state in plain language the main parameters of their automated decisions and how a worker's behaviour affects them. The paper argues such rules are exactly what a large language model can read.

With the rules disclosed, drivers follow the platform more closely. In Manhattan, the share of relocation offers they decline falls from about four in five to one in four under both the reactive and the anticipatory policy.

In Manhattan, disclosure costs the reactive policy R 3.7% of its profit, because R now pays for moves drivers used to make for free, while the anticipatory policy A's profit is unchanged. The two policies, which tie on profit under full compliance, end 6.1% apart.

In the three-zone model, disclosure raises driver income under both policies and both LLMs. For users there, it raises the service rate by 10 to 18 percentage points.

No. Jev appears as one of two LLM driver agents inside a policy-evaluation framework, not as a model being scored for calibration or accuracy. The paper's contribution is a framework for evaluating rebalancing policies under the disclosure requirement, which it says is the first to use LLM agents for this.