Empirical phenomena supporting binding set contraction

Identify empirical phenomena in real-world large language model aggregation that support the binding set contraction mechanism, wherein aggregation combines outputs that are binding with respect to conic feasibility constraints into an output with fewer binding constraints (i.e., an interior point), with particular attention to cases arising from the interaction between prompt-engineering limitations and model capability constraints.

Background

The paper introduces three mechanisms—feasibility expansion, support expansion, and binding set contraction—through which aggregating outputs from multiple copies of the same model can expand the set of elicitable outputs in a principal–agent framework. Binding set contraction occurs when aggregation combines outputs lying on the boundary of conic feasibility constraints into an aggregated output in the interior, thereby reducing the number of binding constraints and making certain outputs elicitably preferable under monotone rewards.

While the authors provide theoretical characterizations and a toy empirical illustration for binding set contraction, they note that broader, real-world empirical evidence for this mechanism remains to be established. Specifically, because binding set contraction emerges from the interplay between prompt-engineering limitations (reward specification over coarser features) and model capability constraints (conic constraints on outputs), identifying concrete empirical settings and phenomena that exhibit this mechanism is left unresolved.