Research map / Multi-agent systems and games / Games and equilibria

Optimizing Social Utility in Sequential Experiments

NeurIPS 2026 · Multi-agent systems and games › Games and equilibria · approach: Strategic classification · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a subsidized sequential RCT approval protocol with anytime-valid e-value tests, solves the agent's belief MDP by dynamic programming, and finds the regulator's optimal subsidy via divide-and-conquer, reporting >35% social utility gains on antibiotic data.

Problem

High costs of large randomized controlled trials deter developers from pursuing uncertain 'moonshot' products with high social utility, such as antibiotics, so regulators need protocols and subsidies that incentivize sequential experimentation.

Approach

Model the developer as a Bayesian agent in a belief MDP over Beta posteriors, with the regulator running an anytime-valid e-value test and partially subsidizing trial costs; show agent utility is piecewise linear and convex in the subsidy, enabling an exact divide-and-conquer search for the socially optimal subsidy.</ismp>

Key results

• Under the optimal subsidy 0.108 for rho_S = $2000M, social utility rises ~5.5% and opt-out probability falls ~22% relative to no subsidy.
• Sequential protocol increases social utility by ~50-60% over non-sequential no-subsidy protocol and by more than 35% over non-sequential optimal-subsidy protocol.
• Partition of the subsidy range has 49 intervals; the algorithm computes the optimal subsidy in a few minutes.

Surprising findings

• The optimal subsidy maximizing anticipated social utility also maximizes true social utility in the experiments.

Limitations

• Beta-Binomial model with a single-arm efficacy test; extensions to dependence or safety are computationally harder.
• Agent benefit rho_A assumed known; no uncertainty in agent benefit.
• Principal assumed to know a prior over agent's prior only through Q; non-economic factors like trial duration are not modeled.

Experimental setup

Datasets: publicly available antibiotic development and approval data (cost and sales estimates)

Benchmarks: simulated antibiotic development setting

Metrics: social utility optimal subsidy agent opt-out probability

Baselines: non-sequential protocol without subsidy non-sequential protocol with optimal subsidy

Builds on: e-values belief Markov decision process Bates et al. contract design for regulatory approval Ville's inequality

Closest papers by meaning

• Strategic Hypothesis Testing (NeurIPS 2025 · similarity 0.83)
• Efficient Adaptive Experimentation with Noncompliance (NeurIPS 2025 · similarity 0.83)
• Steering the Herd: A Framework for LLM-based Control of Social Learning (ICLR 2026 · similarity 0.83)
• Data-driven Design of Randomized Control Trials with Guaranteed Treatment Effects (ICML 2025 · similarity 0.83)
• A Principled Approach to Randomized Selection under Uncertainty: Applications to Peer Review and Grant Funding (NeurIPS 2025 · similarity 0.83)
• Multidimensional Bayesian Utility Maximization: Tight Approximations to Welfare (NeurIPS 2025 · similarity 0.83)