Black-Box Diagnostic for Machine Collectives: Dispersion-Revision Coupling

A new arXiv paper (2608.03722) introduces a black-box diagnostic for evaluating whether LLM collectives genuinely revise their epistemic stance when their outputs become more dispersed. The paper argues that in collective intelligence research, disagreement is often treated as evidence of epistemic diversity, but in LLM collectives, diverse-looking arguments may preserve the same conclusion. The authors operationalize 'dispersion-revision coupling'—the degree to which an intervention that increases output dispersion in embedding space is accompanied by genuine revision rather than premise-preserving reformulation. The diagnostic is black-box, operating solely on generated text, with no assumptions about internal model representations. Two channels are measured independently: an output channel using the Coherence Index (CI) to verify changes in output dispersion, and an epistemic channel using per-turn stance annotation to measure whether the collective revised. The paper is categorized as a new announcement on arXiv, with the abstract available at the provided URL.