Research map / Learning theory and algorithms / Prediction and decision losses

The Limits of AI-Driven Allocation: Optimal Screening under Aleatoric Uncertainty

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Algorithms with explicit predictors · 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.

Studies optimal combination of screening (observing true outcomes) and risk-score-based allocation under a fixed budget, showing the optimal screening band sits at the margin of algorithmic allocation and that screening value is concave in budget.

Problem

How should screening (physical verification) and algorithmic targeting from Bayes-optimal risk scores be combined to allocate a scarce resource under irreducible aleatoric uncertainty and fixed budgets?

Approach

A two-stage framework where a screening stage reveals true outcomes for an interval of risk scores just below the direct-allocation band; closed-form/fixed-point characterization of the optimal screening band, plus a distribution-free concavity result for the value of screening.

Key results

• On synthetic uniform risk with beta=35%, optimal screening at alpha=beta=35% raises precision from 82% to 98%; from 61% to 83% for unimodal; from 50% to 75% for the point mass.
• In Colombia demining (beta=10%), XGBoost precision rises from 39% to 70% at alpha=beta; random screening only reaches 35% to 39%.
• On ACS 2018 low-income prediction (beta=30%), XGBoost precision rises from 65% to 80% at alpha=beta; random screening reaches only 60% to 65%.

Surprising findings

• Random screening yields negligible gains compared with optimal screening placement.
• Screening gains are largest when risk scores are uninformative (point mass at 0.5), even though risk scores carry no targeting information.

Limitations

• Assumes Bayes-optimal risk scores; estimation error is not modeled theoretically and may shift the optimal screening set.
• Allocation and screening budgets treated as exogenous inputs.
• Real-world experiments mix aleatoric uncertainty with estimation error.

Experimental setup

Datasets: Colombia landmine contamination dataset (Granada municipality) ACS 2018 PUMS via folktables synthetic Beta(t,t) risk populations

Benchmarks: synthetic Beta(t,t) populations Colombia landmine contamination ACS 2018 low-income prediction

Metrics: allocation precision utility gap percentage of landmines detected

Baselines: random screening heuristic screening band below algorithmic threshold no-screening algorithmic allocation

Builds on: Perdomo et al. allocation with model improvement vs expanded access proxy means testing

Closest papers by meaning

• The Hidden Cost of Waiting for Accurate Predictions (ICLR 2025 · similarity 0.86)
• Allocation Requires Prediction Only if Inequality Is Low (ICML 2024 · similarity 0.84)
• Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.84)
• On the Meta-Design of Allocation Problems (NeurIPS 2026 · similarity 0.83)
• The Relative Value of Prediction in Algorithmic Decision Making (ICML 2024 · similarity 0.83)
• AI Progress Should Be Measured by Capability-Per-Resource, Not Scale Alone: A Framework for Gradient-Guided Resource Allocation in LLMs (NeurIPS 2025 · similarity 0.83)