ARTFEED — Contemporary Art Intelligence

Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization

ai-technology · 2026-05-09

A research paper introduces Joint Consistency, a test-time aggregation framework formulated as a constrained Ising-type energy minimization problem. It integrates independent evaluation signals as external fields and pairwise comparisons as interactions, subsuming existing voting and weighted aggregation methods. The interaction matrix leverages LLM-as-a-judge comparisons with theoretical interpretation under answer-level homogeneity assumptions. An efficient approximation strategy enables practical large-scale test-time aggregation. The paper is published on arXiv under ID 2605.06219.

Key facts

  • Joint Consistency is a test-time aggregation framework.
  • It is formulated as a constrained Ising-type energy minimization problem.
  • Independent evaluation signals act as external fields.
  • Pairwise comparisons act as interactions.
  • It subsumes existing voting and weighted aggregation methods.
  • The interaction matrix uses LLM-as-a-judge comparisons.
  • It has theoretical interpretation under answer-level homogeneity assumptions.
  • An efficient approximation strategy is developed for large-scale aggregation.

Entities

Institutions

  • arXiv

Sources