Meeting Uncertain Threats with Feedback

What the math gives to ML

The paper gives a constructive framework for allocating a fixed batch of scarce actions across multiple jobs whose completion requirements are uncertain and only revealed through feedback. Its key transferable asset is the combination of a robust fair-allocation baseline with difficulty-aware greedy allocation, together with geometric survival probabilities that quantify the value of another attempt. This maps naturally to inference-time compute scheduling: allocate a fixed number of expert calls, refinement steps, verifier calls, or denoising updates among unresolved examples, then remove examples when feedback indicates completion. The most practical first transfer is a feedback scheduler that falls back to fairness when difficulty estimates are unreliable and uses marginal-success scores when they become informative.